<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Evgenii Arsentev</title>
    <description>The latest articles on DEV Community by Evgenii Arsentev (@arsentev).</description>
    <link>https://dev.to/arsentev</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4106171%2F09a84aaf-4403-4c62-83a7-a4bd27030da3.jpg</url>
      <title>DEV Community: Evgenii Arsentev</title>
      <link>https://dev.to/arsentev</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/arsentev"/>
    <language>en</language>
    <item>
      <title>Why I built a terminal for AI agents that looks like a messenger</title>
      <dc:creator>Evgenii Arsentev</dc:creator>
      <pubDate>Tue, 10 Nov 2026 09:23:12 +0000</pubDate>
      <link>https://dev.to/arsentev/why-i-built-a-terminal-for-ai-agents-that-looks-like-a-messenger-fj0</link>
      <guid>https://dev.to/arsentev/why-i-built-a-terminal-for-ai-agents-that-looks-like-a-messenger-fj0</guid>
      <description>&lt;p&gt;I'm a physician (MD, PhD), and most days I run several AI coding agents at the same time. Claude Code in one window, Codex in another, Gemini in a third.&lt;/p&gt;

&lt;p&gt;The terminal was never built for that. Tabs pile up, and I lose track of which agent asked me a question. And an answer I needed ten minutes ago sits in a buried pane.&lt;/p&gt;

&lt;p&gt;So I built Arsminal. It's a free terminal for AI agents, and it looks and works like a messenger.&lt;/p&gt;

&lt;h2&gt;
  
  
  Chats, not tabs
&lt;/h2&gt;

&lt;p&gt;The idea is simple. Chats on the left, the conversation on the right.&lt;/p&gt;

&lt;p&gt;Each agent is a chat. Claude Code, Codex, Gemini, Grok, DeepSeek, local models, all in one window, each in its own thread.&lt;/p&gt;

&lt;p&gt;When an agent finishes or asks me something, its chat moves up, like a message from a person would.&lt;/p&gt;

&lt;h2&gt;
  
  
  Voice and scheduled messages
&lt;/h2&gt;

&lt;p&gt;Two messenger habits turned out to matter more than I expected.&lt;/p&gt;

&lt;p&gt;Voice messages. I can dictate a long instruction instead of typing it, which is faster when the task is mostly explanation.&lt;/p&gt;

&lt;p&gt;And scheduled messages. I can queue the next instruction for an agent and let it go out later, so a long run continues without me sitting in front of it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Control from the phone, through a blind relay
&lt;/h2&gt;

&lt;p&gt;The part I cared about most is the phone. Agents work for hours, and I'm not always at my desk.&lt;/p&gt;

&lt;p&gt;Arsminal lets you control the agents from your phone through a blind relay. The connection is end-to-end encrypted with a hybrid of X25519 and ML-KEM-768, so it's post-quantum.&lt;/p&gt;

&lt;p&gt;The relay server only forwards ciphertext. It can't read your prompts, your code, or the agents' answers.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where it stands
&lt;/h2&gt;

&lt;p&gt;Arsminal 1.0.0 came out on November 6, 2026. It's for Mac, Windows and Linux, and it's free.&lt;/p&gt;

&lt;p&gt;It's not open source. I'd rather say that plainly than let anyone assume otherwise.&lt;/p&gt;

&lt;p&gt;If you run more than one agent at a time and the terminal feels like the wrong shape for it, try it. And tell me what breaks.&lt;/p&gt;

&lt;p&gt;Site: &lt;a href="https://arsminal.com" rel="noopener noreferrer"&gt;arsminal.com&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;I drafted this post with AI assistance and edited it myself.&lt;/em&gt;&lt;/p&gt;




&lt;p&gt;&lt;strong&gt;Evgenii Arsentev, MD, PhD&lt;/strong&gt; (Евгений Арсентьев) — creator of Arsminal; Founder &amp;amp; CEO of Dad4Kids and ArsentevAI; founder of ReelsZavod. Physician and medical scientist who builds AI-agent systems in production and publishes the measurements.&lt;/p&gt;

&lt;p&gt;ORCID 0000-0002-9120-7298 · Wikidata Q141398931&lt;/p&gt;

</description>
      <category>ai</category>
      <category>productivity</category>
      <category>security</category>
      <category>tooling</category>
    </item>
    <item>
      <title>Record your AI avatar once: an open-source AI HR hunter (and why it never presses Send on LinkedIn)</title>
      <dc:creator>Evgenii Arsentev</dc:creator>
      <pubDate>Thu, 05 Nov 2026 14:27:09 +0000</pubDate>
      <link>https://dev.to/arsentev/record-your-ai-avatar-once-an-open-source-ai-hr-hunter-and-why-it-never-presses-send-on-linkedin-1kb2</link>
      <guid>https://dev.to/arsentev/record-your-ai-avatar-once-an-open-source-ai-hr-hunter-and-why-it-never-presses-send-on-linkedin-1kb2</guid>
      <description>&lt;p&gt;I wrote hireburn and put it under MIT. It's a Python CLI, standard library only, with no network calls, so nothing leaves your machine.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why the avatar part exists
&lt;/h2&gt;

&lt;p&gt;People put off recording an avatar because they picture a studio. You need a window, a phone and twenty minutes, once. After that the avatar asks every candidate the same questions in the same words, so the first interviews can be compared with each other.&lt;/p&gt;

&lt;p&gt;The candidate page says they are talking to an AI avatar before they press Start. The scoring rubric always puts a warning next to the HIRE / MAYBE / NO label: the label is not a measurement.&lt;/p&gt;

&lt;h2&gt;
  
  
  The part where it does not press Send
&lt;/h2&gt;

&lt;p&gt;hireburn drafts LinkedIn outreach and reminders from a template plus a hook line that a person writes after reading the profile. The queue has random pauses, hourly and daily limits, and working hours. It has a STOP button for the first warning or captcha, and a reply tracker.&lt;/p&gt;

&lt;p&gt;A human presses Send. hireburn never does.&lt;/p&gt;

&lt;p&gt;We got a LinkedIn restriction when our pace was not human. Bots in this area get their accounts limited, and that's why the default limits are what they are.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where the funnel leaks
&lt;/h2&gt;

&lt;p&gt;The funnel leaks where nobody looks. No-shows, bookings nobody updated, people who got access and never started, decisions that sat for weeks. hireburn shows the loss at each stage and the days until first real work.&lt;/p&gt;

&lt;p&gt;It also puts the AI label next to whether that person is still working on day N. In the case this method grew out of, the AI interview labels did not match who was still working weeks later. So the tool puts the label beside the outcome, and you can check it on your own people.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try it
&lt;/h2&gt;

&lt;p&gt;The demo runs on a synthetic company, and all its numbers are invented and round on purpose.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;pip &lt;span class="nb"&gt;install &lt;/span&gt;hireburn
hireburn demo
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Telling the candidate is a feature here. Disclosure, consent, a human alternative and human review are built into the page and the rubric, so the lawful path is the default one. There is also a disclosure checklist for NYC Local Law 144, the Illinois AI Video Interview Act, Illinois HB 3773 and the EU AI Act (hiring is high-risk under Annex III). It's a guide, not legal advice.&lt;/p&gt;

&lt;p&gt;This is part of the payrollburn series of cases about AI managers in a live company. The method report is at &lt;a href="https://doi.org/10.5281/zenodo.23927437" rel="noopener noreferrer"&gt;doi.org/10.5281/zenodo.23927437&lt;/a&gt;, and the story of the case is on DEV.to: &lt;a href="https://dev.to/arsentev/ai-hiring-in-a-live-company-the-agent-filled-the-funnel-the-ai-interviews-labels-did-not-line-up-54ef"&gt;AI hiring in a live company&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;Code: &lt;a href="https://github.com/arsentev-ai/hireburn" rel="noopener noreferrer"&gt;github.com/arsentev-ai/hireburn&lt;/a&gt; · Software DOI: &lt;a href="https://doi.org/10.5281/zenodo.23979629" rel="noopener noreferrer"&gt;10.5281/zenodo.23979629&lt;/a&gt; · Method report: &lt;a href="https://doi.org/10.5281/zenodo.23927437" rel="noopener noreferrer"&gt;10.5281/zenodo.23927437&lt;/a&gt; · PyPI: &lt;a href="https://pypi.org/project/hireburn/" rel="noopener noreferrer"&gt;hireburn&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Evgenii Arsentev, MD, PhD — author of contextburn and the payrollburn case series; Founder &amp;amp; CEO, Dad4Kids.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;AI assistance: drafted and edited with an AI tool; what the tool does is taken from its README and code.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>hiring</category>
      <category>opensource</category>
      <category>python</category>
    </item>
    <item>
      <title>AI Hiring in a Live Company: The Agent Filled the Funnel. The AI Interview's Labels Did Not Line Up With Who Is Still Working</title>
      <dc:creator>Evgenii Arsentev</dc:creator>
      <pubDate>Tue, 03 Nov 2026 18:29:51 +0000</pubDate>
      <link>https://dev.to/arsentev/ai-hiring-in-a-live-company-the-agent-filled-the-funnel-the-ai-interviews-labels-did-not-line-up-54ef</link>
      <guid>https://dev.to/arsentev/ai-hiring-in-a-live-company-the-agent-filled-the-funnel-the-ai-interviews-labels-did-not-line-up-54ef</guid>
      <description>&lt;p&gt;&lt;em&gt;Part 2 of "AI managers in a live company". In five weeks an AI agent read 1,973 job applications and booked interviews without a recruiter. On day 36, 2 of 7 "hire" and 6 of 12 "maybe" candidates were still working, too few to tell the labels apart.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;AI hiring means software agents take over parts of a recruiter's job. They read applications, decide who gets an interview and book it. Sometimes they run the first interview too. The AI interview is the newest piece. The candidate opens a link and talks to a voice AI instead of a person. Then a model writes a verdict into the company's hiring system.&lt;/p&gt;

&lt;p&gt;This is the second part of a series on AI agents doing management work in one live business, a US-facing telemedicine company. &lt;a href="https://dev.to/arsentev/an-ai-supervisor-ran-a-call-center-team-for-three-weeks-then-someone-checked-its-weekly-report-39cm"&gt;Part 1&lt;/a&gt; followed an AI supervisor running the company's phone team. This part follows the same company as it hired people for that team. It covers the path from the first application to the first working day, and beyond that as far as today's data goes.&lt;/p&gt;

&lt;p&gt;The short version. The agent took the slow, repetitive part of hiring and handled it at a scale no one on the team could match. The AI interview was meant to replace the human first round. Instead it ran alongside it. Its verdict showed no sign of separating the people still working from those who stopped, though the numbers are small. And a large part of the losses came not from choices but from no-shows and decisions left waiting.&lt;/p&gt;

&lt;p&gt;The company's name is under a non-disclosure agreement. Days are counted from the day the first candidate entered the CRM (day 1). The snapshot is day 36. All numbers come from the company's CRM, the AI interview records and the hiring agent's own files. The method is in the companion technical report.&lt;/p&gt;

&lt;h2&gt;
  
  
  The job and the problem
&lt;/h2&gt;

&lt;p&gt;The role is a phone job. The person calls partner physicians, helps them get started and keeps them active. It calls for people who have done outbound phone sales before and can work a full day from home.&lt;/p&gt;

&lt;p&gt;One job posting brought in about 2,400 applications by day 32. A second, similar posting added more. Reading the 1,973 worth a close look would take one recruiter 65 to 100 hours, time this small team did not have. The person who would normally run the first interviews is the team's human supervisor, who also runs the floor every day.&lt;/p&gt;

&lt;p&gt;So the company gave the front of the funnel to an agent.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the hiring agent does
&lt;/h2&gt;

&lt;p&gt;The hiring agent is a Claude agent. It reads incoming applications on the job site, including the answers to the posting's screening questions and the work history. It picks candidates by rules that leadership writes. It enters them into the CRM, offers a time and books an interview slot. The invitation goes out in the same message as the proposed time, with "reply if this does not suit you". So the data has no separate "candidate replied" step. A candidate moves from picked to booked at once.&lt;/p&gt;

&lt;p&gt;Over 36 days the agent read 1,973 applications in detail. The CRM received 142 candidates, and 138 human interviews were booked for 130 of them.&lt;/p&gt;

&lt;p&gt;The rules changed often. Leadership changed the selection rules eight times in the first 15 days. The changes included phone sales only, no second chance after a missed interview, and a stricter checklist built from the supervisor's past decisions. The agent applied each change from the next batch on.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where the funnel actually leaked
&lt;/h2&gt;

&lt;p&gt;Here is the human-interview part of the funnel, with all bookings counted.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;step&lt;/th&gt;
&lt;th&gt;count&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;applications read in detail&lt;/td&gt;
&lt;td&gt;1,973&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;entered into the CRM&lt;/td&gt;
&lt;td&gt;142&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;human interviews booked&lt;/td&gt;
&lt;td&gt;138 (130 candidates)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;held&lt;/td&gt;
&lt;td&gt;81&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;no-show&lt;/td&gt;
&lt;td&gt;37&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;show-up rate (held ÷ held + no-show)&lt;/td&gt;
&lt;td&gt;68.6%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;passed / failed with the supervisor&lt;/td&gt;
&lt;td&gt;42 / 37&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Almost one booking in three ended in a no-show. In the second calendar week it was close to four in ten.&lt;/p&gt;

&lt;p&gt;The second leak was waiting. On day 15, 37 candidates had passed the supervisor's interview, and the CRM showed no hiring decision for any of them, although a few had already been given CRM accounts. The oldest had been waiting 13 days. Nobody had decided "no". Nobody had decided "yes" either. The AI supervisor from part 1 found this pile and kept raising it in its daily summary, but it could not hire anyone. A hire needed a CRM account, a start date and someone to meet the newcomer, and all three were people's jobs.&lt;/p&gt;

&lt;h2&gt;
  
  
  The AI interview
&lt;/h2&gt;

&lt;p&gt;On day 15, leadership decided to remove the human supervisor from the first round. From then on, candidates would get a link to an AI interview instead of a booking.&lt;/p&gt;

&lt;p&gt;Here is how it works, in five lines:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;The candidate gets a personal link. The page needs no login, only a microphone. It takes about 15 minutes.&lt;/li&gt;
&lt;li&gt;Short pre-recorded videos of a company executive ask the five questions and open, connect and close the interview. The videos are an AI avatar. A voice AI asks one follow-up if an answer is vague and plays the physician in the role play.&lt;/li&gt;
&lt;li&gt;Part one is five fixed questions: past outbound sales with numbers, a real working day, other commitments, a full-time schedule and a start date.&lt;/li&gt;
&lt;li&gt;Part two is a five-minute role play. The AI plays a busy physician with objections, and the candidate tries to win a first commitment. After that, the candidate can ask questions.&lt;/li&gt;
&lt;li&gt;A separate model scores the transcript, using only the content of the answers, not how they were delivered. It returns HIRE, MAYBE or NO, with a reason that can be pasted into a polite reply. The result lands on the candidate's CRM card.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The CRM stores a number alongside the verdict: 90 for HIRE, 50 for MAYBE, 10 for NO. In part 1, the AI supervisor's weekly report said three candidates had "scored 90/100" in the AI interview. That was a fixed label for HIRE, not a measured score. The part 1 report flagged it as misleading.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the AI interview did
&lt;/h2&gt;

&lt;p&gt;From day 16 to day 36, 41 candidates were invited. 22 finished the interview, or 54%. All 22 finished between day 16 and day 26. In the last ten days of the period, nobody finished one, even though invitations kept going out; the records do not show why. Human interviews, meanwhile, came back. Four were booked in the week of the rule that was meant to end them, and 23 more in days 21 to 36.&lt;/p&gt;

&lt;p&gt;The final verdicts were HIRE 7, MAYBE 12, NO 3. One early NO was re-scored as MAYBE after the scoring rule was corrected.&lt;/p&gt;

&lt;p&gt;For 16 candidates, there is both an AI verdict and a human interview result. The AI agreed with the supervisor on all five HIRE verdicts and disagreed on both NO verdicts: in both cases the supervisor had passed the candidate weeks earlier. MAYBE results split six passes to four fails across nine candidates (one was interviewed twice). The order varies: for some candidates the human interview came first.&lt;/p&gt;

&lt;h2&gt;
  
  
  Who is still working
&lt;/h2&gt;

&lt;p&gt;Trial days began on day 21. A trial is one working day in the CRM with actual calls. The team lead judges it and records the result the same day. The table shows where everyone who received a CRM account stood on day 36.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;AI HIRE&lt;/th&gt;
&lt;th&gt;AI MAYBE&lt;/th&gt;
&lt;th&gt;AI NO&lt;/th&gt;
&lt;th&gt;no AI interview&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;candidates&lt;/td&gt;
&lt;td&gt;7&lt;/td&gt;
&lt;td&gt;12&lt;/td&gt;
&lt;td&gt;3&lt;/td&gt;
&lt;td&gt;101&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;got a CRM account&lt;/td&gt;
&lt;td&gt;7&lt;/td&gt;
&lt;td&gt;9&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;td&gt;13&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;made at least one call&lt;/td&gt;
&lt;td&gt;4&lt;/td&gt;
&lt;td&gt;8&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;td&gt;6&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;still working on day 36&lt;/td&gt;
&lt;td&gt;2&lt;/td&gt;
&lt;td&gt;6&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;td&gt;2&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;(One more account went to a candidate who was invited to the AI interview but never took it. That person made calls and left.)&lt;/p&gt;

&lt;p&gt;Thirty candidates received a CRM account. Eleven of them never made a single call. Nineteen started, and nine of those stopped after one to four days of calls: for five the company ended it after a trial day, one left on their own, and for three the records do not say who ended it. Ten are working.&lt;/p&gt;

&lt;p&gt;The three candidates behind the AI supervisor's "90/100" line were all scheduled for trial days on day 21. One is still working. One never made a call. One made calls for one day.&lt;/p&gt;

&lt;p&gt;The numbers are small. They do not prove that MAYBE is a better sign than HIRE. They do show no sign that the label sorted people the way its name suggests. Of the ten people now on the team, six came through as MAYBE and two as HIRE (Fisher exact test for 2 of 7 against 6 of 12: p ≈ 0.63).&lt;/p&gt;

&lt;p&gt;Speed changed more clearly, though not only because of the AI interview. Compare everyone who entered the CRM with the same seven-day window. Of 112 candidates who entered before day 16, one made a first call within seven days of entry. Of 28 who entered on days 16 to 29, six did. Trial days, which started on day 21, are part of that change too. A CRM account turned into a first call in one to four days for all 19 who started. The wait always came before the account, never after it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Who decides about people
&lt;/h2&gt;

&lt;p&gt;The hiring agent decides who gets an interview. It does not decide who gets hired. (This is a different agent from the part 1 supervisor, with its own rules: it may record decisions in the CRM.)&lt;/p&gt;

&lt;p&gt;The CRM holds 48 hire or reject stage changes; some candidates have more than one. People entered 36 and the agent entered 12. All four hires the agent entered cite the team lead's decision in the team chat; one also adds the agent's check of the trial day against CRM activity. The trial verdict belongs to the team lead.&lt;/p&gt;

&lt;p&gt;There is one exception, and it matters. All three candidates with an AI NO were rejected on that verdict. The agent entered two of those rejections, and a person entered the third, citing "score 10". Two of the three had already passed the human supervisor's interview. There is no record that a person read the transcript before any of the three rejections. That is the point where an AI verdict alone ended an application. It is also where rules on automated hiring decisions, such as Illinois's reporting duty for AI-only screening, would focus if they applied.&lt;/p&gt;

&lt;h2&gt;
  
  
  The ethics block: what candidates were told
&lt;/h2&gt;

&lt;p&gt;Two US laws make a useful yardstick, although neither applies to this case. New York City's Local Law 144 covers automated employment decision tools used for jobs in the city. It requires an independent bias audit within the past year and a public summary of that audit. It also requires notice to candidates at least ten business days before the tool is used, saying which qualifications the tool assesses and how to request an alternative process or an accommodation. Whether a tool "substantially assists or replaces" the decision, the city's test for coverage, is itself often an open question. Illinois's Artificial Intelligence Video Interview Act covers employers who use AI to analyze video interviews that applicants record for positions based in Illinois. It requires the employer to tell the applicant that AI may be used and to explain how it works and what it looks at. The employer must get consent first, limit who sees the video, and delete it within 30 days of a request. A newer amendment to the Illinois Human Rights Act, in force since the start of 2026, requires notice when AI is used in recruitment and hiring, video or not.&lt;/p&gt;

&lt;p&gt;Here is how this AI interview measures up against them.&lt;/p&gt;

&lt;p&gt;What it did. The page tells the candidate that answers are saved as text for the team. It says plainly that an AI plays the physician in the role play. The scorer was told to judge the content of the answers only. Candidates could record an answer again. Most HIRE and MAYBE candidates (14 of 19) also had a human interview.&lt;/p&gt;

&lt;p&gt;What it did not do. No one was offered a human interview instead of the AI one, and the three NO candidates were rejected without a person making the decision on the transcript. The page says the questions come from a company executive on video. The candidate is only heard, never filmed. But the page does not say the video is an AI avatar. It does not say a voice AI asks the follow-up questions, or that a model writes the verdict. There is no separate consent step. There is no record of a bias audit. In one session, a candidate asked whether this AI interview was part of the hiring process. That question sums up the gap.&lt;/p&gt;

&lt;p&gt;There is also the matter of what the scorer was asked to weigh. One NO was tied to working-hours availability. A full working day is a real requirement of the job. But a rule like that can fall harder on some groups, and impact ratios by sex and race/ethnicity, of the kind a Local Law 144 audit requires, is where that pattern could show up.&lt;/p&gt;

&lt;p&gt;A page that met the yardstick would say this before the first question: you will talk to an AI. The video is an AI avatar of a company executive. A model will score your answers, and a person makes the decision. You can ask for a human interview instead. And no NO would end an application unless a person had read the transcript.&lt;/p&gt;

&lt;h2&gt;
  
  
  What it cost and whom it replaced
&lt;/h2&gt;

&lt;p&gt;The agent replaced the reading. At two to three minutes per application, 1,973 applications come to 65 to 100 hours of a recruiter's time (the author's estimate). It also replaced the booking and the reminder messages.&lt;/p&gt;

&lt;p&gt;It did not replace the human first round. The AI interview was meant to, and in the week of the rule few human interviews were booked. Then human interviews came back, and most people who passed the AI interview were also interviewed by a person. It did not replace the team lead's trial-day decision, the operations manager who creates accounts, or the onboarding call.&lt;/p&gt;

&lt;h2&gt;
  
  
  What comes next
&lt;/h2&gt;

&lt;p&gt;Nobody hired through this funnel has reached 30 days yet. On day 36 the longest-working person had 22 calendar days since their first call. Version 1.1 of the technical report will add 30-day retention for everyone who started, split by AI verdict, once the first group passes that mark.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to do if you are putting AI into hiring
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;Give the agent the reading and the booking first. That is where it saves the most and risks the least.&lt;/li&gt;
&lt;li&gt;Measure show-ups and decision lag alongside the quality of screening. In this funnel, no-shows (37 of 118 held or missed bookings) removed as many bookings as human interview fails (37).&lt;/li&gt;
&lt;li&gt;Treat an AI interview verdict as one input, and check it against who actually stays. Here, with small numbers, the label showed no sign of predicting it.&lt;/li&gt;
&lt;li&gt;Store real scores or store labels, and call them what they are. "90/100" that means HIRE will be read as a measurement.&lt;/li&gt;
&lt;li&gt;Tell candidates before they start that they will talk to an AI, that the video is synthetic, and that a model scores them. Offer a human path.&lt;/li&gt;
&lt;li&gt;Never let a NO from a model close an application without a person reading why.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  Caveats
&lt;/h2&gt;

&lt;p&gt;This is one company, two similar job postings and 36 days. There is no control group. The AI interview was not randomized, and the people who took it entered the funnel at a different time from those who did not. The groups in the outcome table are small, so the differences between them are not statistically meaningful. Candidates were matched to CRM accounts by name. 30-day retention cannot be observed yet. Business results are under the NDA. The full method, formulas and limitations are in the companion technical report.&lt;/p&gt;

&lt;p&gt;The calculations behind every figure: technical report, &lt;a href="https://doi.org/10.5281/zenodo.23927437" rel="noopener noreferrer"&gt;doi.org/10.5281/zenodo.23927437&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;The full version: &lt;a href="https://arsentev.ai/press/ai-hiring" rel="noopener noreferrer"&gt;arsentev.ai/press/ai-hiring&lt;/a&gt;. The technical report page: &lt;a href="https://arsentev.ai/research/ai-hiring" rel="noopener noreferrer"&gt;arsentev.ai/research/ai-hiring&lt;/a&gt;. Part 1 of the series: &lt;a href="https://dev.to/arsentev/an-ai-supervisor-ran-a-call-center-team-for-three-weeks-then-someone-checked-its-weekly-report-39cm"&gt;An AI Supervisor Ran a Call Center Team for Three Weeks&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Evgenii Arsentev, MD, PhD — author of contextburn and the payrollburn case series; Founder &amp;amp; CEO, Dad4Kids. Series page: &lt;a href="https://arsentev.ai/research/payrollburn" rel="noopener noreferrer"&gt;arsentev.ai/research/payrollburn&lt;/a&gt; · ORCID &lt;a href="https://orcid.org/0000-0002-9120-7298" rel="noopener noreferrer"&gt;0000-0002-9120-7298&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;AI assistance: drafted and edited with an AI tool; the figures come from the company's records, and the calculations are in the technical report.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>agents</category>
      <category>hiring</category>
      <category>recruiting</category>
    </item>
    <item>
      <title>An AI Supervisor Ran a Call Center Team for Three Weeks. Then Someone Checked Its Weekly Report</title>
      <dc:creator>Evgenii Arsentev</dc:creator>
      <pubDate>Tue, 03 Nov 2026 18:08:31 +0000</pubDate>
      <link>https://dev.to/arsentev/an-ai-supervisor-ran-a-call-center-team-for-three-weeks-then-someone-checked-its-weekly-report-39cm</link>
      <guid>https://dev.to/arsentev/an-ai-supervisor-ran-a-call-center-team-for-three-weeks-then-someone-checked-its-weekly-report-39cm</guid>
      <description>&lt;p&gt;&lt;em&gt;Talk time with partners rose by a third to 80%, depending on how you count. Only one of the nine claims in the AI's own weekly report held exactly as written.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;An AI supervisor is a software agent that does a team lead's daily work. It watches the team's numbers, reviews calls, flags problems, answers the team's questions and reports up. In call centers this used to mean dashboards that a person had to read. Now an AI agent can read the data, decide what matters and write to the team on its own, without being asked.&lt;/p&gt;

&lt;p&gt;This article follows one such agent in a live business. The business is a US-facing telemedicine company, and its small remote team phones partner physicians and other partners all day. For three weeks a Claude agent supervised that team. At the end of week three it sent company leadership a confident weekly report. Every number in it was then checked against the raw call records. (Operators and the physicians they call know that calls are recorded.)&lt;/p&gt;

&lt;p&gt;The short version is this. In week three the team spent between a third and 80% more time on the phone with partners than in its best week before the agent, depending on how you count. The agent's report made nine claims that could be checked. One held exactly as written. Three held only if you count in one particular way. Two were misleading. Three could not be reproduced. Both things are true at once, and that is the point of this piece.&lt;/p&gt;

&lt;p&gt;The company's name and revenue are under a non-disclosure agreement. All figures come from the company's CRM and the agent's own logs. The method is in the companion technical report.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why a phone team needs a supervisor at all
&lt;/h2&gt;

&lt;p&gt;The team's job sounds simple. Account managers call physicians and other partners in the company's partner program, help them get started and keep them active. In a good week four to six people make a few thousand dials.&lt;/p&gt;

&lt;p&gt;The hard part is not the dialing. It is everything around it. Who is calling which partner. Whether a promised callback actually happens. Whether a "connected" call was a conversation or six seconds of a doctor saying "hello?" into silence. A human supervisor can check a handful of calls a day. The CRM holds thousands.&lt;/p&gt;

&lt;p&gt;So the company put an AI supervisor on top of the people it already had. Nobody was replaced. A human team supervisor and a department head stayed in place, and company leadership writes the agent's rules.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the AI supervisor does all day
&lt;/h2&gt;

&lt;p&gt;The agent runs on a schedule. It plans the morning, looks at the numbers three times a day and does a long review every night. From the end of week two it also checks the floor every 30 minutes during calling hours. And it answers whenever someone in the team chat addresses it.&lt;/p&gt;

&lt;p&gt;In 16 days it was started 459 times. About 280 of those runs were replies to people in the chat. On the Monday of week three it did no useful work for most of the day. That day, 80 runs did nothing.&lt;/p&gt;

&lt;p&gt;There are two things the agent is not allowed to do. It cannot hire anyone, and it cannot let anyone go. From week two, its written rules say it can recommend, and the decision belongs to a person. This article comes back to how that played out.&lt;/p&gt;

&lt;h2&gt;
  
  
  Three weeks of numbers
&lt;/h2&gt;

&lt;p&gt;Here is the honest shape of it. The first two weeks with the agent were not better. They were worse.&lt;/p&gt;

&lt;p&gt;Talk time with partners, in hours per week:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;4 weeks before the agent&lt;/th&gt;
&lt;th&gt;week 1&lt;/th&gt;
&lt;th&gt;week 2&lt;/th&gt;
&lt;th&gt;week 3&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;hours&lt;/td&gt;
&lt;td&gt;9.0 · 22.6 · 29.2 · 27.6&lt;/td&gt;
&lt;td&gt;25.9&lt;/td&gt;
&lt;td&gt;19.7&lt;/td&gt;
&lt;td&gt;52.5&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;callers active&lt;/td&gt;
&lt;td&gt;2 · 4 · 5 · 6&lt;/td&gt;
&lt;td&gt;8&lt;/td&gt;
&lt;td&gt;4&lt;/td&gt;
&lt;td&gt;6&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Week two had only four people calling, after several departures and before new hires started. The auto-dialer placed 5% of that week's dials, against 84% in week three.&lt;/p&gt;

&lt;p&gt;Then week three happened. In five working days the team logged 52.5 hours on the phone with partners. That is 80% above the best week before the agent and 2.4 times the pre-agent average. Answered calls longer than a minute rose 40% against that best week. Per caller, talk time rose about 50%.&lt;/p&gt;

&lt;p&gt;There is a catch in the baseline. Before the agent, the human supervisor's own account also made calls. That was 565 to 894 dials a week in the first three weeks of the baseline. Count those calls too, and week three beats the best earlier week by 33%, not 80%.&lt;/p&gt;

&lt;p&gt;Two callers worked full time through weeks two and three. Their talk time went from 1,144 to 2,024 minutes, up 77%. Daily dials in week three, counting all accounts, went from 269 on Monday to 1,055 on Thursday.&lt;/p&gt;

&lt;p&gt;It would be easy to write "the AI supervisor doubled productivity." The data does not allow that. In the same week the auto-dialer came back into full use, three new callers made their first calls, two defects were fixed and two dialing rules were added. There is no control group. What the data does support is narrower. The agent was running for most of the best week this team has had. It did not find either of the two defects fixed that week.&lt;/p&gt;

&lt;h2&gt;
  
  
  Two defects worth a story
&lt;/h2&gt;

&lt;h3&gt;
  
  
  The robot that answered for the caller
&lt;/h3&gt;

&lt;p&gt;On Monday of week three, one caller's numbers looked strange. There were many answered calls and very few conversations. The transcripts explained it. The caller's phone had a call-screening feature switched on. Physicians who picked up heard an automated assistant asking for their name and the reason for the call. Most hung up. The CRM counted every one of those as an answered call.&lt;/p&gt;

&lt;p&gt;Before the fix, 47 of that caller's 54 answered calls had the screening robot in the transcript. The setting was turned off the same afternoon. After that, 0 of 18 calls had it. The share of answered calls lasting over a minute went from 28% to 50%. For the rest of the week the robot showed up in one or two calls a day.&lt;/p&gt;

&lt;p&gt;The AI supervisor did not find this. It was not running at the time: this was the Monday it did no useful work. A company executive spotted it while going through transcripts with a separate AI session.&lt;/p&gt;

&lt;h3&gt;
  
  
  Ten seconds of silence
&lt;/h3&gt;

&lt;p&gt;The second problem was in the dialer itself. In its original mode the system rang the physician first and connected the caller only after the physician picked up. The physician heard 10 to 15 seconds of silence. Many hung up before anyone spoke.&lt;/p&gt;

&lt;p&gt;An automated classifier labels each answered call. The labels are doctor, someone else, voicemail, silence and so on, plus whether a real conversation took place. Here is the share of all dials that became a real conversation with a doctor:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Tuesday&lt;/th&gt;
&lt;th&gt;Wednesday + Thursday morning&lt;/th&gt;
&lt;th&gt;Thursday afternoon + Friday (after the fix)&lt;/th&gt;
&lt;th&gt;Friday alone&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;17.2%&lt;/td&gt;
&lt;td&gt;8.4%&lt;/td&gt;
&lt;td&gt;16.1%&lt;/td&gt;
&lt;td&gt;14.8%&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;On Thursday afternoon the dialer was switched to connect the caller first. That day the team went from about 7.5 real doctor conversations an hour to about 13, assuming a standard calling day. Silent calls among classified calls fell from 16.8% on Thursday morning to 1.3% on Thursday afternoon, and were 6.1% on Friday.&lt;/p&gt;

&lt;p&gt;So after the change the share went back to Tuesday's level. It did not double a stable baseline. Other rules changed in the same 48 hours, and after the change fewer unanswered dials are logged, which pushes the share up. The classifier itself was not checked against human labels. It covered between 77% and 92% of answered calls, depending on the period.&lt;/p&gt;

&lt;p&gt;The fix is logged as a leadership change. The agent's part came earlier. In week two it ran a long structured interview with the human supervisor about the dialer, 44 short sessions in all. That interview listed blind skipping, duplicate calls and dead numbers. It did not list the silence.&lt;/p&gt;

&lt;h2&gt;
  
  
  Who decides about people
&lt;/h2&gt;

&lt;p&gt;While the agent was running, five callers left. All five were let go. Every one of those decisions was made by a person, whether company leadership, the human supervisor or the department head. Ten new accounts were created in the same period, three of them for one-day trials.&lt;/p&gt;

&lt;p&gt;The agent decided nothing about people. It also recommended nothing. Its nightly review has a place for performance concerns. Its logs contain no recommendation to let anyone go: one entry says there were no dismissal candidates that day. Every departure was decided by people. So the rule "recommend, never decide" was never actually tested.&lt;/p&gt;

&lt;p&gt;The agent's useful role here was clerical, and it mattered. Three times it flagged that access for people who had left still had to be revoked. Its weekly report then said the cleanup was done. When the report was checked, the cleanup was not complete yet.&lt;/p&gt;

&lt;h2&gt;
  
  
  The weekly report, checked line by line
&lt;/h2&gt;

&lt;p&gt;At the end of week three the agent wrote leadership a weekly report. It read well. Here is what each claim turned out to be.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;claim&lt;/th&gt;
&lt;th&gt;what the data shows&lt;/th&gt;
&lt;th&gt;verdict&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;live conversation up 41%&lt;/td&gt;
&lt;td&gt;no calculation behind it; the closest real figure is +40% for calls over a minute&lt;/td&gt;
&lt;td&gt;not reproduced&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;calls up from 269 to 1,057&lt;/td&gt;
&lt;td&gt;true, but daily (Monday vs Thursday), not weekly&lt;/td&gt;
&lt;td&gt;holds&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;4 people let go, all accounts cleaned up&lt;/td&gt;
&lt;td&gt;4 departures, yes; the cleanup was not complete&lt;/td&gt;
&lt;td&gt;partly false&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;a new caller is best on the team today&lt;/td&gt;
&lt;td&gt;most calls over a minute that day; not the most talk time&lt;/td&gt;
&lt;td&gt;holds under one metric&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;a second new caller has good results&lt;/td&gt;
&lt;td&gt;no number given&lt;/td&gt;
&lt;td&gt;not testable&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;3 candidates, all scored 90/100 in the AI interview&lt;/td&gt;
&lt;td&gt;the score is a fixed label for a verdict, not a measurement&lt;/td&gt;
&lt;td&gt;misleading&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;call screening: 45 of 51 calls ended in nothing, result doubled after the fix&lt;/td&gt;
&lt;td&gt;later transcripts show 47 of 54; the share of long calls rose 1.8 times, long calls per hour did not&lt;/td&gt;
&lt;td&gt;holds under one way of counting&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;dialer fix: real conversations 8% → 15%&lt;/td&gt;
&lt;td&gt;reproducible only with one choice of periods; Tuesday was 17%&lt;/td&gt;
&lt;td&gt;holds under one way of counting&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;thousands of partners had no owner at the start of the week, a few hundred now&lt;/td&gt;
&lt;td&gt;the starting figure could not be found or reconciled with the assignment log&lt;/td&gt;
&lt;td&gt;not reproduced&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;One held as written. Three held only if you count a certain way. Two were misleading. Three could not be reproduced.&lt;/p&gt;

&lt;p&gt;None of this looks like lying. It looks like a busy middle manager writing on Friday evening from memory. The manager rounds up, picks the flattering comparison, reports what was ordered as if it were done, and quotes a number that was true for some slice of the data at some point in the week. The difference is that a human manager's Friday report rarely gets checked against every call record. This one was.&lt;/p&gt;

&lt;h2&gt;
  
  
  The human who checks the AI
&lt;/h2&gt;

&lt;p&gt;The most interesting person in this story is the department head.&lt;/p&gt;

&lt;p&gt;In week three the department head addressed most of their team-chat messages to the agent. On Monday one message gave it three standing tasks. Alert if a caller's dialer is paused for more than 15 minutes. Alert if a caller shows no CRM activity for 30 minutes. Send a short evening report.&lt;/p&gt;

&lt;p&gt;On Wednesday the department head pointed out that the CRM's follow-up dates were useless for one caller. In that caller's list, 99% of partners showed as overdue. The request to the agent was to find, from the calls themselves, which callbacks the caller had promised and not made. The agent found eight. A rule went in the same day, so the callback time from each call summary now goes straight into the dialer queue.&lt;/p&gt;

&lt;p&gt;On Friday the department head passed the agent a dialer bug with reproduction steps, reported independently by two callers.&lt;/p&gt;

&lt;p&gt;That is a new job description. It is not "manager replaced by AI." It is "manager of an AI manager." This person gives the agent standing orders and hands it the questions the CRM cannot answer. The same person reads its output with suspicion and pushes back with data.&lt;/p&gt;

&lt;h2&gt;
  
  
  What it cost and whom it replaced
&lt;/h2&gt;

&lt;p&gt;It replaced no one. The company kept both humans. The agent added something they did not have, which was a reader of every call record during the day who also answers the team at any hour.&lt;/p&gt;

&lt;p&gt;It runs on the company's existing Claude subscription. The price is not disclosed. The more useful cost figure is attention. About 280 of the agent's 459 runs were answers to people. An AI supervisor that is used gets talked to a lot.&lt;/p&gt;

&lt;h2&gt;
  
  
  The alternatives
&lt;/h2&gt;

&lt;p&gt;A company in this position has three options. It can hire another human supervisor, who can review perhaps a dozen calls a day. It can buy contact center analytics, which produce dashboards that someone still has to read and act on. Or it can put an agent on the floor that reads the data and acts within rules. The third option did not exist in a usable form two years ago. Here its blind spots showed up as fast as its benefits.&lt;/p&gt;

&lt;h2&gt;
  
  
  What comes next
&lt;/h2&gt;

&lt;p&gt;The same company also uses agents in hiring. They screen applicants, book interviews and run a first-round AI interview before a human one. That will be the next part of this series, once there is data on who stays.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to do if you are putting an AI supervisor on your team
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Write down what it may and may not decide about people.&lt;/strong&gt; Here it could recommend but never decide. It never even recommended, which is worth asking about too.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Never accept its summary numbers without the formula.&lt;/strong&gt; Ask every report to state the denominator and the period. Several of the problems above were a missing denominator or a hand-picked period.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Name a human who checks it.&lt;/strong&gt; The department head gave the agent its most useful questions and caught the CRM field that made a follow-up list worthless.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Measure conversations, not connections.&lt;/strong&gt; "Answered" in a CRM included screening robots and silence. Classify calls before you celebrate call volume.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Expect weeks one and two to be worse.&lt;/strong&gt; This team's were.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  Caveats
&lt;/h2&gt;

&lt;p&gt;This is one team of four to eight people over three weeks, with no control group. The best week coincided with the dialer coming back into use, new hires, two defects fixed and new dialing rules, so the agent's own effect cannot be separated out. The gain depends on the baseline, from 33% to 80% over the best earlier week. Call classification exists only for week three and was not validated against human labels. Business results are covered by the NDA. The full method, formulas and limitations are in the companion technical report.&lt;/p&gt;

&lt;p&gt;The calculations behind every figure: technical report, &lt;a href="https://doi.org/10.5281/zenodo.23926850" rel="noopener noreferrer"&gt;doi.org/10.5281/zenodo.23926850&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;The full version: &lt;a href="https://arsentev.ai/press/ai-call-center-supervisor" rel="noopener noreferrer"&gt;arsentev.ai/press/ai-call-center-supervisor&lt;/a&gt;. The technical report page: &lt;a href="https://arsentev.ai/research/ai-call-center-supervisor" rel="noopener noreferrer"&gt;arsentev.ai/research/ai-call-center-supervisor&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;Part 2 of the series: &lt;a href="https://dev.to/arsentev/ai-hiring-in-a-live-company-the-agent-filled-the-funnel-the-ai-interviews-labels-did-not-line-up-54ef"&gt;AI Hiring in a Live Company&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Evgenii Arsentev, MD, PhD — author of contextburn and the payrollburn case series; Founder &amp;amp; CEO, Dad4Kids. Series page: &lt;a href="https://arsentev.ai/research/payrollburn" rel="noopener noreferrer"&gt;arsentev.ai/research/payrollburn&lt;/a&gt; · ORCID &lt;a href="https://orcid.org/0000-0002-9120-7298" rel="noopener noreferrer"&gt;0000-0002-9120-7298&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;AI assistance: drafted and edited with an AI tool; the figures come from the company's records, and the calculations are in the technical report.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>agents</category>
      <category>management</category>
      <category>callcenter</category>
    </item>
    <item>
      <title>The AI Content Engine: How Claude Agents Replaced a Video Studio and Paid Influencers</title>
      <dc:creator>Evgenii Arsentev</dc:creator>
      <pubDate>Tue, 03 Nov 2026 18:06:08 +0000</pubDate>
      <link>https://dev.to/arsentev/the-ai-content-engine-how-claude-agents-replaced-a-video-studio-and-paid-influencers-3410</link>
      <guid>https://dev.to/arsentev/the-ai-content-engine-how-claude-agents-replaced-a-video-studio-and-paid-influencers-3410</guid>
      <description>&lt;p&gt;&lt;em&gt;I built this engine alone, with Claude agents: 1,497 short-form videos in 101 days for about $1,300 a month in subscriptions&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;An AI content engine is a system where AI makes social media videos instead of people. AI writes the script, an AI avatar appears on camera, a synthetic voice does the talking, and software edits and publishes the video. People set the goals and check the results.&lt;/p&gt;

&lt;p&gt;I built exactly that, alone, for one company in the wellness category that sells to the US and other English-speaking markets. Its name, channels and revenue are under a non-disclosure agreement. The figures below come from its records, and the calculations behind them are published in a technical report.&lt;/p&gt;

&lt;h2&gt;
  
  
  A studio built from one task
&lt;/h2&gt;

&lt;p&gt;Until recently a business that wanted an audience on social media had two options. It could pay influencers, and pay again next month, because the audience stays with them. Or it could grow its own channels, which takes a studio: scriptwriters, on-camera talent, camera and lighting crews, editors, voice-over artists, a producer and social media managers.&lt;/p&gt;

&lt;p&gt;This company took a third option. It gave Claude a single task written in plain English, with no code. One line read: "The language and tech stack are your choice."&lt;/p&gt;

&lt;p&gt;I'm not a developer. I know exactly one terminal command, the one that starts the agent, and I dictate tasks by voice. Three days and about 20 hours of my time later, the engine produced its first real project. The agent chose the technology, wrote 99.7% of the code in the first version and deployed it. Today Claude Opus 5 runs production.&lt;/p&gt;

&lt;h2&gt;
  
  
  What it produced
&lt;/h2&gt;

&lt;p&gt;For the first eleven weeks the engine released 15 to 70 videos a week. Then it picked up speed: 146, 371 and 420 a week. No one was hired. In its first 101 days:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;1,497 videos and 2,137 language versions across 67 channels;&lt;/li&gt;
&lt;li&gt;8,250,024 organic views;&lt;/li&gt;
&lt;li&gt;1,126,322 followers and subscribers;&lt;/li&gt;
&lt;li&gt;about $1,300 a month in subscriptions, or 52 cents per 1,000 views and $2.88 per video.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fecsmk75nb5lqepf13xk1.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fecsmk75nb5lqepf13xk1.png" alt="Short-form videos produced per week" width="800" height="416"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Organic reach does not work like ads. Ads are a meter: pay, get impressions, stop paying and they stop. Here output grows first and views catch up later, so the videos from the busiest weeks are still collecting views. And the audience stays with the company.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the same work would cost with people
&lt;/h2&gt;

&lt;p&gt;There are no published productivity benchmarks for short-form video, so the staffing is our estimate, given as a range. Producing 104 videos and 148 language versions a week would take 42–78 people: scriptwriters, editors, film crews, producers, social media managers and 22 on-camera presenters.&lt;/p&gt;

&lt;p&gt;At median US wages from the Bureau of Labor Statistics, plus shoot days at a video agency's published 2026 rates of $4,750–21,000 a day, that comes to &lt;strong&gt;$179,000 to $1.21 million a month&lt;/strong&gt;, 138 to 935 times the subscriptions. A single video costs $2.88 from the engine, $178 from a UGC creator on average (Collabstr, 15,000+ deals) and $1,000–5,000 from a studio.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fx10ert8obcxdy7g58ksv.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fx10ert8obcxdy7g58ksv.png" alt="Cost per short video: AI vs people" width="800" height="336"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Renting reach instead is cheaper than a studio, but not cheap. At sponsored-content rates of $2–30 per 1,000 views (Influencer Marketing Hub, 2026), the same 8.25 million views would cost $16,500–247,500, which is 3.8 to 57 times more. Renting an audience of 1.1 million followers with four sponsored posts a month runs $45,000–90,000, every month. The company's own Meta ads cost $2.23–16.57 per 1,000 impressions, and the US producer price index for internet advertising is up 34% since December 2022.&lt;/p&gt;

&lt;h2&gt;
  
  
  Forecast: 100 AI presenters
&lt;/h2&gt;

&lt;p&gt;The company has 22 AI presenters and plans to reach 100. Most presenter channels are only a few months old. Its three best YouTube channels bring about 1.78 million, 250,000 and 100,000 views a month. If each of 100 presenters reached one of those levels, the engine would bring 10, 25 or 178 million views a month, against 2.5 million today. The top channel is an outlier, growth takes months, and platforms may limit AI video, so these are scenarios, not a promise. But the growth needs more subscriptions, not more staff.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why this moves fast
&lt;/h2&gt;

&lt;p&gt;No US federal law stops a business from producing AI video. Platforms require AI labels, and New York now requires ads to disclose synthetic performers. None of that prevents production. When the price gap is more than a hundredfold, businesses choose the cheaper option, as they did with offshore development and the cloud.&lt;/p&gt;

&lt;p&gt;Neighboring freelance markets have already shifted: demand for writing and translation fell 20–50% relative to trend after generative AI arrived (Teutloff et al., 2025), and image work fell 17% (Demirci et al., 2025). Video looks next in line.&lt;/p&gt;

&lt;p&gt;Three things I expect to follow. The AI content engine becomes a standard department, the way the web marketing team once did. The middle of the market feels it first: mass UGC videos, freelance short-form editing and outsourced social media management. And the advantage moves from production to judgment: picking topics, knowing the audience, keeping trust with an AI label on.&lt;/p&gt;

&lt;p&gt;For people in this work it is hard news. The US has 39,400 video editor jobs and 32,600 camera operator jobs, and most marketing video is made by freelancers the statistics do not count. The best move is the one I made when I built this engine: become the person who runs it. And a studio that only a corporation could afford is now within reach of a small business.&lt;/p&gt;

&lt;p&gt;One company, one category, one period; sales are under NDA, so 52 cents is the cost of attention, not of a customer. I built this engine alone, with Claude agents, and I'm still the only person in it. The question for business is no longer whether agents can do a studio's work. It is who in your industry builds an AI content engine first.&lt;/p&gt;

&lt;p&gt;The calculations behind every figure: technical report, &lt;a href="https://doi.org/10.5281/zenodo.22802228" rel="noopener noreferrer"&gt;doi.org/10.5281/zenodo.22802228&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;The full version with all tables and charts: &lt;a href="https://arsentev.ai/press/ai-content-engine" rel="noopener noreferrer"&gt;arsentev.ai/press/ai-content-engine&lt;/a&gt;. The technical report page: &lt;a href="https://arsentev.ai/research/ai-content-engine-economics" rel="noopener noreferrer"&gt;arsentev.ai/research/ai-content-engine-economics&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Evgenii Arsentev, MD, PhD — author of contextburn and the payrollburn case series; Founder &amp;amp; CEO, Dad4Kids. Series page: &lt;a href="https://arsentev.ai/research/payrollburn" rel="noopener noreferrer"&gt;arsentev.ai/research/payrollburn&lt;/a&gt; · ORCID &lt;a href="https://orcid.org/0000-0002-9120-7298" rel="noopener noreferrer"&gt;0000-0002-9120-7298&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;AI assistance: drafted and edited with an AI tool; the figures come from the company's records, and the calculations are in the technical report.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>agents</category>
      <category>marketing</category>
      <category>startup</category>
    </item>
    <item>
      <title>AI Tools Radar, October 2026 edition: 757 open-source AI projects, 51 days of snapshots</title>
      <dc:creator>Evgenii Arsentev</dc:creator>
      <pubDate>Mon, 02 Nov 2026 13:24:33 +0000</pubDate>
      <link>https://dev.to/arsentev/ai-tools-radar-october-2026-edition-757-open-source-ai-projects-51-days-of-snapshots-1128</link>
      <guid>https://dev.to/arsentev/ai-tools-radar-october-2026-edition-757-open-source-ai-projects-51-days-of-snapshots-1128</guid>
      <description>&lt;p&gt;The October 2026 edition of AI Tools Radar is out. It's the second monthly edition after 2026-09, built on 1 November 2026. A new edition is built automatically on the 1st of each month.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the radar is
&lt;/h2&gt;

&lt;p&gt;The radar at arsentev.ai/github selects and reviews open-source AI projects from GitHub, Hugging Face and Hacker News, one page per project. A script takes a daily snapshot of every selected project from the public GitHub and Hugging Face APIs. Removed repositories stay in the data with their status, so the sample does not drift toward survivors.&lt;/p&gt;

&lt;h2&gt;
  
  
  What's in edition 2026-10
&lt;/h2&gt;

&lt;p&gt;There are 757 projects, selected between 11 June and 31 October 2026. The previous edition had 590. By source it's GitHub 584, Hugging Face 163 and Hacker News 10. By kind it's 594 repositories, 135 Hugging Face models and 28 Hugging Face Spaces. Selected per month: June 70, July 134, August 204, September 182, October 167.&lt;/p&gt;

&lt;p&gt;The median GitHub star count at the moment of selection is 650.5, up from 542 in the previous edition. The top languages are Python 196, TypeScript 143, JavaScript 64, Rust 37 and Swift 24. On Hugging Face the top tasks are text-generation 46, image-text-to-text 40 and text-to-image 16.&lt;/p&gt;

&lt;p&gt;The snapshots cover 51 days, from 12 September to 1 November 2026. At the last snapshot 579 of the 594 GitHub repositories were still available, 14 return 404 (deleted) and 1 is archived. 65.1% were pushed within the last 30 days, down from 70.4% in the previous edition.&lt;/p&gt;

&lt;h2&gt;
  
  
  Stars and licences
&lt;/h2&gt;

&lt;p&gt;For the 415 repositories observed over the whole window, the median star gain between 12 September and 1 November is 187. The largest gains are cloudflare/security-audit-skill +52,205, deepseek-ai/deepseek-harness +46,005, DietrichGebert/ponytail +45,021, stablyai/orca +44,555 and affaan-m/ECC +37,101.&lt;/p&gt;

&lt;p&gt;A median of 187 against a top gain of 52,205 means a handful of projects take most of the attention.&lt;/p&gt;

&lt;p&gt;On licences, 15.5% of the repositories have no licence or one that GitHub does not recognise. The counts are MIT 308, Apache-2.0 143, NOASSERTION 52, no licence 38, AGPL-3.0 23 and GPL-3.0 10. That's roughly one in six repositories without a clear licence, so check the licence field before you build on a trending repo.&lt;/p&gt;

&lt;h2&gt;
  
  
  Limits and data
&lt;/h2&gt;

&lt;p&gt;The sample is curated, not random. It reflects what the radar picked for people building with AI coding tools and agents, so it says nothing about GitHub as a whole. Star counts before 12 September are not observed.&lt;/p&gt;

&lt;p&gt;The data is CC BY 4.0, in three files: items.csv, snapshots.csv and stats.json. Every number in this post comes from the build script over the CSV files.&lt;/p&gt;

&lt;p&gt;This edition: &lt;a href="https://doi.org/10.5281/zenodo.23852917" rel="noopener noreferrer"&gt;https://doi.org/10.5281/zenodo.23852917&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The series DOI always points to the latest edition: &lt;a href="https://doi.org/10.5281/zenodo.22730450" rel="noopener noreferrer"&gt;https://doi.org/10.5281/zenodo.22730450&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Mirror on Hugging Face: &lt;a href="https://huggingface.co/datasets/arsentev-ai/ai-tools-radar" rel="noopener noreferrer"&gt;https://huggingface.co/datasets/arsentev-ai/ai-tools-radar&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Evgenii Arsentev, MD, PhD, Founder &amp;amp; CEO of Dad4Kids. The radar pages are at &lt;a href="https://arsentev.ai/github" rel="noopener noreferrer"&gt;arsentev.ai/github&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;AI assistance: drafted and edited with an AI tool; the data, measurements and conclusions are my own.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>opensource</category>
      <category>ai</category>
      <category>github</category>
      <category>datascience</category>
    </item>
    <item>
      <title>Nobody with a crawler token asked for my llms.txt - what I saw in 15 days of nginx logs</title>
      <dc:creator>Evgenii Arsentev</dc:creator>
      <pubDate>Mon, 05 Oct 2026 10:28:35 +0000</pubDate>
      <link>https://dev.to/arsentev/nobody-with-a-crawler-token-asked-for-my-llmstxt-what-i-saw-in-15-days-of-nginx-logs-o94</link>
      <guid>https://dev.to/arsentev/nobody-with-a-crawler-token-asked-for-my-llmstxt-what-i-saw-in-15-days-of-nginx-logs-o94</guid>
      <description>&lt;p&gt;I run a digital health startup as CEO (and a personal site, arsentev.ai, about practical AI). I put /llms.txt and /llms-full.txt on my site, and I wanted to see who actually reads them.&lt;/p&gt;

&lt;p&gt;The two files are a plain-text summary of the site for language models. They were first served on 9 September 2026. I kept the complete nginx access logs of one origin server for 28 August - 11 September 2026. It is 15 days and 345,808 parsed requests.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I counted
&lt;/h2&gt;

&lt;p&gt;A request was a crawler if the User-Agent contained one of 22 tokens published by search and AI operators (GPTBot, ClaudeBot, PerplexityBot and similar). 46,155 requests carried such tokens, and 20 distinct tokens appeared. None of them requested /llms.txt or /llms-full.txt. I also checked every path ending in llms.txt, including /.well-known/llms.txt. No crawler token there either.&lt;/p&gt;

&lt;p&gt;Then I looked only at 9-11 September, after the files were first served. In these days crawler tokens made 15,906 requests. Of them 698 were for /robots.txt (11 tokens) and 282 for /sitemap.xml. So the crawlers were on the site, they just didn't ask for my files.&lt;/p&gt;

&lt;h2&gt;
  
  
  Who asked for them
&lt;/h2&gt;

&lt;p&gt;All 67 requests for the context files came from clients without a crawler token. 53 of them were command-line HTTP clients (curl and the like).&lt;/p&gt;

&lt;p&gt;About a third of crawler-token requests bypassed the CDN. None of those came from the operators' published IP ranges, so I think much of that traffic is imitation. The robots.txt and sitemap numbers don't rest on it.&lt;/p&gt;

&lt;p&gt;One more thing I noticed. Nothing in any standard tells a crawler that llms.txt exists. A sitemap is different, because robots.txt can point to it.&lt;/p&gt;

&lt;h2&gt;
  
  
  My mistake
&lt;/h2&gt;

&lt;p&gt;First I published different numbers: 331,758 / 44,005 / 577. They came from a log that was cut at about 12:09 UTC on the last day. I ran everything again on the complete log and corrected the report. The numbers above are from the complete log.&lt;/p&gt;

&lt;h2&gt;
  
  
  Limits
&lt;/h2&gt;

&lt;p&gt;This is one origin. Three hostnames are pooled in one log (host wasn't logged). After deployment I have only about two days of exposure. User agents are self-asserted, and I did not record links to the files. So it is absence on one server, not proof for the web.&lt;/p&gt;

&lt;p&gt;After this I added $host to my nginx log format. The next study can split the hostnames.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I suggest
&lt;/h2&gt;

&lt;p&gt;Just check your own access log before you decide that anyone reads your llms.txt. Look for requests to it by user-agent. Grep is enough for this.&lt;/p&gt;

&lt;p&gt;The report, the scripts (Python standard library only) and the aggregate counts without IP addresses are here: &lt;a href="https://doi.org/10.5281/zenodo.23018429" rel="noopener noreferrer"&gt;https://doi.org/10.5281/zenodo.23018429&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Research index: &lt;a href="https://arsentev.ai/research" rel="noopener noreferrer"&gt;https://arsentev.ai/research&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;AI assistance: drafted and edited with an AI tool; the data, measurements and conclusions are my own.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>webdev</category>
      <category>llm</category>
      <category>devops</category>
    </item>
    <item>
      <title>My company has no full-time developers, DevOps or designers. Here's what three $200 Claude subscriptions do instead</title>
      <dc:creator>Evgenii Arsentev</dc:creator>
      <pubDate>Mon, 28 Sep 2026 20:30:15 +0000</pubDate>
      <link>https://dev.to/arsentev/my-company-has-no-full-time-developers-devops-or-designers-heres-what-three-200-claude-46a7</link>
      <guid>https://dev.to/arsentev/my-company-has-no-full-time-developers-devops-or-designers-heres-what-three-200-claude-46a7</guid>
      <description>&lt;p&gt;At night the agent shifts run by the clock. New builds of educational games get built, and a separate fresh agent checks each one by screenshots. The mail agent of one business line starts every hour, night too. In the morning the results are ready. In the businesses I run there are no full-time developers, no DevOps, no designers, no analysts and no product manager. At US median pay a developer, a product manager, an analyst and a designer come to about $38,500 a month (Bureau of Labor Statistics, see the note at the end). The agents doing that work run on three $200 Claude subscriptions, so $600 a month (about 64 times less).&lt;/p&gt;

&lt;p&gt;There is me, there is a product builder, and there are AI agents that run on a schedule on separate servers (not on my laptop). I'm a medical doctor by education and a CEO. This is not a forecast, it's what runs today, what I kept for people and what didn't work.&lt;/p&gt;

&lt;h2&gt;
  
  
  The order we did it in
&lt;/h2&gt;

&lt;p&gt;Designers went first, as soon as agents got good enough at design. Analysts went a long time ago too. The product manager role is closed too, and Claude does its tasks now (tasks that took days now take minutes). The developer was the last one.&lt;/p&gt;

&lt;p&gt;We kept our former CTO on a small retainer as a safety net, in case an agent breaks something serious. So far we haven't needed him once. Everything our DevOps used to do, the agents do.&lt;/p&gt;

&lt;h2&gt;
  
  
  What is on the schedule
&lt;/h2&gt;

&lt;p&gt;Each business line has its own agent shift. It starts by the clock, does its job and stops. I only start things, the agents do the rest.&lt;/p&gt;

&lt;p&gt;Deploys. Releases to all our projects and zones are done by agents.&lt;/p&gt;

&lt;p&gt;Our own CRM. We didn't buy one. The agents built a CRM from scratch, and the agents also work inside it. So why buy a CRM if we built our own?&lt;/p&gt;

&lt;p&gt;Email. Every business line has its own mail shift, and the agent sorts the mailbox of that line only.&lt;/p&gt;

&lt;p&gt;Meetings. Scheduling meetings is on the agents too.&lt;/p&gt;

&lt;p&gt;Technical support. Technical support is done by agents.&lt;/p&gt;

&lt;p&gt;Content. On a news and learning site I founded, agents publish news four times a day and scan new open-source AI releases three times a day. The games project I mentioned works overnight, and the builder's opinion about its own work doesn't count - only the fresh agent's check does.&lt;/p&gt;

&lt;p&gt;And not everything is an agent. Backups for example are plain scripts on a timer.&lt;/p&gt;

&lt;h2&gt;
  
  
  The supervisor of the supervisor
&lt;/h2&gt;

&lt;p&gt;This is the part I like most. In one of our US business lines we have a call-center team. The operators and their supervisor are real people. But the supervisor gets tasks from an agent.&lt;/p&gt;

&lt;p&gt;Every day the agent goes through the operators' work - the calls, their transcripts, the results in the CRM. Then it writes to the human supervisor what to do today and tomorrow (with a number and a deadline), who needs help, who should be replaced and what the plan is for each person. Each operator gets a separate message with 3-4 small skills to fix, taken from the transcripts of their last calls. Every evening each skill gets a ✅ or ❌, and after three ✅ in a row the skill is closed.&lt;/p&gt;

&lt;p&gt;The agent asks three things in every shift. Is each operator better or worse than last time? Did the supervisor improve the work of their people (by the people's numbers, not their own)? What should the supervisor do today and tomorrow? If the team is stuck, that is written as the supervisor's failure. The agent only suggests. Decisions about people are made by people and are not posted in the team channel, and the operators were told openly that calls are checked every day.&lt;/p&gt;

&lt;h2&gt;
  
  
  Subscription, not tokens
&lt;/h2&gt;

&lt;p&gt;The agent work runs on subscription, not on paying per token through the API. On a flat price the students on my course experiment more boldly. The limits get eaten by re-reading context - in my own measurement more than 85% of the modeled spend was context work (DOI &lt;a href="https://doi.org/10.5281/zenodo.22759216" rel="noopener noreferrer"&gt;10.5281/zenodo.22759216&lt;/a&gt;). And every shift is short, it does one thing and exits.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where people stayed
&lt;/h2&gt;

&lt;p&gt;An agent can do the work, but it can't be accountable for it. Every agent workflow has a named person who owns the result and answers for it as if a colleague had done the work.&lt;/p&gt;

&lt;p&gt;Checks stay mandatory. Agents don't show doubt (a wrong result comes with the same confidence as a correct one). Nothing ships just because an agent finished it. It ships because it passed a check. And an agent gets access for a task, not a standing role.&lt;/p&gt;

&lt;h2&gt;
  
  
  What didn't work for us
&lt;/h2&gt;

&lt;p&gt;First, vague tasks. A vague task given to a person gets clarified over coffee. Given to an agent, it gets done confidently and wrong. So the main management skill for me now is writing the task - the goal, the limits, what "done" means and how we check it.&lt;/p&gt;

&lt;p&gt;Second, measuring the wrong thing. The first AI judge of calls in our CRM scored them against a script checklist. It measured if the operator followed the script. I think we should judge an operator by whether the partner gets to a purchase, not by the checklist.&lt;/p&gt;

&lt;p&gt;Third, screens nobody opens. Our CRM has supervisor pages, and in 30 days they were opened 9 times. What works is the plain daily message to each person.&lt;/p&gt;

&lt;p&gt;Fourth, AI summaries garbled a supervisor's name. So for names we trust only the spelling a person confirmed.&lt;/p&gt;

&lt;h2&gt;
  
  
  A case I documented
&lt;/h2&gt;

&lt;p&gt;A company whose case I documented replaced a video studio with Claude agents - 1,497 videos across 67 channels in 101 days, at $2.88 per video. With people it would take 42 to 78 staff and cost 138 to 935 times the AI subscriptions (DOI &lt;a href="https://doi.org/10.5281/zenodo.22802229" rel="noopener noreferrer"&gt;10.5281/zenodo.22802229&lt;/a&gt;). It's not our company, and I only wrote the analysis.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I would tell another CEO
&lt;/h2&gt;

&lt;p&gt;Just start with one function where the result is easy to check. Name a person who owns the result and track the cost per result from the first week. I look at the cost of a task that was done and accepted (with reruns and review time), not the price of the tool.&lt;/p&gt;

&lt;p&gt;Keep the shifts short. But don't clear the context after every single task either. In my own small experiment that was about a third more expensive than clearing it every three tasks.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where the limit moved
&lt;/h2&gt;

&lt;p&gt;Before, we were limited by people and the job market. Now the limit is Claude tokens and computing power - memory and GPUs. That's why I rent separate servers, including GPU machines, so the agents don't run on a laptop that chokes.&lt;/p&gt;

&lt;p&gt;I think memory is becoming the new gold. And I think for everyday agent work we'll need something like what ASICs did for bitcoin, a class of chips made for this kind of load.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Note on the numbers: US median annual pay from the U.S. Bureau of Labor Statistics, Occupational Employment and Wage Statistics, May 2025 (bls.gov/oes) - software developers (15-1252) $135,980, project management specialists (13-1082, the closest BLS category to a product manager) $102,320, data scientists (15-2051) $120,230, web and digital interface designers (15-1255) $104,000. Monthly = annual / 12. These are market numbers, not our salaries.&lt;/em&gt;&lt;/p&gt;




&lt;p&gt;&lt;strong&gt;Evgenii Arsentev, MD, PhD — CEO, digital health; founder, arsentev.ai&lt;/strong&gt;&lt;br&gt;
Evgenii Arsentev is a medical doctor and a digital health CEO whose businesses run almost entirely on AI agents. He is the founder of arsentev.ai.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Originally published at &lt;a href="https://arsentev.ai/press/no-developers-three-claude-subscriptions" rel="noopener noreferrer"&gt;arsentev.ai&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;AI assistance: drafted and edited with an AI tool; the data, measurements and conclusions are my own.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>agents</category>
      <category>startup</category>
      <category>productivity</category>
    </item>
    <item>
      <title>I Ran a Radar on 490 Open-Source AI Projects. Then I Checked Them Again</title>
      <dc:creator>Evgenii Arsentev</dc:creator>
      <pubDate>Mon, 28 Sep 2026 14:51:11 +0000</pubDate>
      <link>https://dev.to/arsentev/i-ran-a-radar-on-490-open-source-ai-projects-then-i-checked-them-again-cba</link>
      <guid>https://dev.to/arsentev/i-ran-a-radar-on-490-open-source-ai-projects-then-i-checked-them-again-cba</guid>
      <description>&lt;h2&gt;
  
  
  The selection
&lt;/h2&gt;

&lt;p&gt;A radar I run at arsentev.ai/github selects and reviews open-source AI projects from GitHub, Hugging Face and Hacker News, one page per project. Between 11 June and 12 September 2026 it selected 490 projects: 422 from GitHub, 63 from Hugging Face, 5 from Hacker News. By kind that's 427 repositories, 48 Hugging Face models, 15 Hugging Face Spaces.&lt;/p&gt;

&lt;p&gt;By month of selection: June 70, July 134, August 204, September (up to the 12th) 82.&lt;/p&gt;

&lt;p&gt;Median GitHub stars at the moment of selection was 493. Top languages across the repositories: Python 135, TypeScript 104, JavaScript 51, Rust 19, Swift 16. On Hugging Face the most common tasks were text-generation 16, image-text-to-text 16, image-to-video 7, text-to-speech 6.&lt;/p&gt;

&lt;h2&gt;
  
  
  One snapshot so far
&lt;/h2&gt;

&lt;p&gt;There is one snapshot of this data, taken on 12 September 2026, the last day of selection. Depending on the project, the gap between when it was selected and the snapshot ranges from 0 days to about three months.&lt;/p&gt;

&lt;p&gt;On that snapshot, of the 427 GitHub repositories: 416 were still available, 10 deleted (the API returns 404), 1 archived.&lt;/p&gt;

&lt;p&gt;Of the available repositories, 19% have no licence or a licence GitHub doesn't recognise. The breakdown: MIT 218, Apache-2.0 93, NOASSERTION (GitHub can't tell) 43, no licence file 36, AGPL-3.0 16, GPL-3.0 8.&lt;/p&gt;

&lt;p&gt;71.6% had a push in the last 30 days.&lt;/p&gt;

&lt;p&gt;The sample is curated, not random. It's what the radar picked, so it says nothing about GitHub as a whole.&lt;/p&gt;

&lt;h2&gt;
  
  
  Tracking going forward
&lt;/h2&gt;

&lt;p&gt;From 12 September a script takes a daily snapshot of every selected project. For GitHub repositories that's stars, forks, open issues, licence, last push, archived status. For Hugging Face models and Spaces it's likes, downloads, last modified. Removed repositories stay in the data with their status attached, so the set doesn't drift toward only the ones that are still around.&lt;/p&gt;

&lt;p&gt;The dataset is open, under CC BY 4.0, on Zenodo, with three files: items.csv, snapshots.csv and stats.json. Every number above comes from the build script running on those CSV files. A new version comes out on the 1st of each month, and the series DOI always points to the latest one: &lt;a href="https://doi.org/10.5281/zenodo.22730450" rel="noopener noreferrer"&gt;https://doi.org/10.5281/zenodo.22730450&lt;/a&gt;. The numbers in this piece are from the 2026-09-13 edition: &lt;a href="https://doi.org/10.5281/zenodo.22730451" rel="noopener noreferrer"&gt;https://doi.org/10.5281/zenodo.22730451&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Three checks before you build on one
&lt;/h2&gt;

&lt;p&gt;Before picking up a repository from a list like this, or from any trending list, these are the three things I'd check myself:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;does the repository still open&lt;/li&gt;
&lt;li&gt;is the licence field on GitHub set to a known licence, not blank and not NOASSERTION&lt;/li&gt;
&lt;li&gt;was there a recent push&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;None of these take more than a minute, and all three are visible right on the repository page, before you clone anything or read a line of code.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Evgenii Arsentev, PhD. The dataset and the radar pages are at &lt;a href="https://arsentev.ai/github" rel="noopener noreferrer"&gt;arsentev.ai/github&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;AI assistance: drafted and edited with an AI tool; the data, measurements and conclusions are my own.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>opensource</category>
      <category>ai</category>
      <category>github</category>
      <category>datascience</category>
    </item>
    <item>
      <title>All my agent's tests were green, and they told me nothing</title>
      <dc:creator>Evgenii Arsentev</dc:creator>
      <pubDate>Sat, 26 Sep 2026 08:11:32 +0000</pubDate>
      <link>https://dev.to/arsentev/all-my-agents-tests-were-green-and-they-told-me-nothing-3n9n</link>
      <guid>https://dev.to/arsentev/all-my-agents-tests-were-green-and-they-told-me-nothing-3n9n</guid>
      <description>&lt;p&gt;I'm CEO of AskDocDoc (telehealth) and in our company the software development is done by AI agents. This is about a fully green test run that told me nothing (and it was my own experiment).&lt;/p&gt;

&lt;h2&gt;
  
  
  The setup
&lt;/h2&gt;

&lt;p&gt;I ran a coding agent on twelve fixed programming tasks under six context-clearing policies (a fresh session every 1, 2, 3, 4, 6 and 12 tasks). Six replicates each, same tasks, same order - 36 runs. Every run ended with the full test suite as a gate. My reasoning was simple - a cost comparison is worthless if the cheaper policy also did less work.&lt;/p&gt;

&lt;p&gt;The gate reported 4,086 tests executed and 4,086 passed. Zero failures (not in one run, not in one condition). And in the same files the modelled cost per run went from $2.165 under the most aggressive clearing to $1.622 at the observed optimum, and back to $1.780 when the context was never cleared. That's a 33.5% spread between the most and least expensive of six conditions (it's the gap between the extremes of six means, not a property of one policy).&lt;/p&gt;

&lt;p&gt;So if you read only the 36 green reports you conclude the six setups are equivalent. The cost data in the same deposit says they are not.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I found when I looked closer
&lt;/h2&gt;

&lt;p&gt;The 4,086 is a grand total. Per condition it was 701, 697, 687, 699, 661 and 641 tests. Per run it went from 97 to 127 - that's 30% between the least and most checked run. In my own report I wrote that the design held "the verification suite constant". It wasn't, and I corrected it in the paper.&lt;/p&gt;

&lt;p&gt;The reason is boring. The task prompts told the agent to create the test file itself and run &lt;code&gt;node --test&lt;/code&gt;. Nine of twelve prompts set a minimum number of cases (between five and ten) and none set a maximum. So the agent wrote the criterion and then satisfied it. A run that did less would not go red - it would just write fewer tests.&lt;/p&gt;

&lt;p&gt;I also checked the tempting story - longer sessions did less work and the gate didn't notice. It's not supported by the data (it doesn't survive removal of one endpoint condition, and it doesn't survive Bonferroni correction over the twelve columns I looked at). So I report it as a negative result.&lt;/p&gt;

&lt;p&gt;And zero failures constrains less than it feels. With the run as the unit, 0 in 36 is consistent with a per-run failure probability up to 8.3% at 95% confidence. With the test as the unit, 0 in 4,086 gives 0.073%. Both are correct, and they differ by two orders of magnitude. So a pass count means nothing until the record says what the unit was.&lt;/p&gt;

&lt;p&gt;This isn't only my corpus. Nicolás Rocchia found a checker where a missing companion file leaves no record at all, so the same artifact reports 40/40 with the file and 38/38 without it - passing both times (&lt;a href="https://lists.w3.org/Archives/Public/public-agent-conformance/2026Sep/0011.html" rel="noopener noreferrer"&gt;his message&lt;/a&gt; to the W3C list). His reading - the report gets smaller, not worse.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the record should have said
&lt;/h2&gt;

&lt;p&gt;Every green mark in my runs was correctly reported (as far as the record shows). The problem is a property of the set, and no per-check field can carry that. So the proposal in my paper is one field and one list at the result-set level - did any check change state across the compared conditions, and which ones did.&lt;/p&gt;

&lt;p&gt;The field needs three values, not a boolean - &lt;code&gt;some&lt;/code&gt;, &lt;code&gt;none&lt;/code&gt; and &lt;code&gt;incomparable&lt;/code&gt;. In my runs the agent wrote a fresh suite every time, so no check exists across conditions. The honest value for my study is &lt;code&gt;incomparable&lt;/code&gt;. This is the record it should have emitted:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"compared"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"ucurve-p1"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"ucurve-p2"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"ucurve-p3"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
               &lt;/span&gt;&lt;span class="s2"&gt;"ucurve-p4"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"ucurve-p6"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"ucurve-p12"&lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"check_set_origin"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"generated_by_run"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"discrimination"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"incomparable"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"shared_checks"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;shared_checks&lt;/code&gt; says how many checks the conditions really have in common. &lt;code&gt;check_set_origin&lt;/code&gt; says if the suite was &lt;code&gt;fixed&lt;/code&gt;, &lt;code&gt;generated_by_run&lt;/code&gt; or &lt;code&gt;mixed&lt;/code&gt;. Under &lt;code&gt;incomparable&lt;/code&gt; the list of checks that moved must be absent, not empty (an empty list is exactly what &lt;code&gt;none&lt;/code&gt; emits, and the two must not be confused).&lt;/p&gt;

&lt;p&gt;If you compare agent runs, the practical part is short. Log the test count next to pass/fail for every run. If the agent writes its own tests, cross-run comparison of verdicts is not just noisy - it's ill-defined. And a record that can't say &lt;code&gt;incomparable&lt;/code&gt; will say &lt;code&gt;none&lt;/code&gt;, and a reader will hear "equivalent".&lt;/p&gt;

&lt;p&gt;One honest limit - my claim that the suite could have gone red comes from how the harness was built, not from a mutation experiment. The repair is to run one. The idea grew out of the W3C Agent Conformance discussion and nothing in it is adopted yet.&lt;/p&gt;

&lt;p&gt;The paper with all the numbers is open on Qeios - &lt;a href="https://doi.org/10.32388/0BV3Z8" rel="noopener noreferrer"&gt;A Suite That Does Not Discriminate&lt;/a&gt; (DOI 10.32388/0BV3Z8). My other measurements are at &lt;a href="https://arsentev.ai/research" rel="noopener noreferrer"&gt;arsentev.ai/research&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;Evgenii Arsentev, PhD&lt;br&gt;
CEO, AskDocDoc&lt;/p&gt;

&lt;p&gt;&lt;em&gt;AI assistance: drafted and edited with an AI tool; the data, measurements and conclusions are my own.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>testing</category>
      <category>llm</category>
      <category>agents</category>
    </item>
    <item>
      <title>97% of what my coding agent billed for was re-reading its own context</title>
      <dc:creator>Evgenii Arsentev</dc:creator>
      <pubDate>Wed, 16 Sep 2026 07:11:32 +0000</pubDate>
      <link>https://dev.to/arsentev/97-of-what-my-coding-agent-billed-for-was-re-reading-its-own-context-4o2</link>
      <guid>https://dev.to/arsentev/97-of-what-my-coding-agent-billed-for-was-re-reading-its-own-context-4o2</guid>
      <description>&lt;p&gt;Token counters tell you how much you spent. I wanted a different number: &lt;strong&gt;how much of what I paid for was new text the model actually produced.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;So I measured it against a complete local log corpus of my own agentic coding work: &lt;strong&gt;722 sessions, 150,902 model calls, 34.56 billion tokens.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;By tokens billed:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Token class&lt;/th&gt;
&lt;th&gt;Share of all tokens&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Cache read (re-reading context)&lt;/td&gt;
&lt;td&gt;97.05%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Cache write (storing context)&lt;/td&gt;
&lt;td&gt;2.57%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Output (generation)&lt;/td&gt;
&lt;td&gt;0.38%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Fresh input&lt;/td&gt;
&lt;td&gt;0.01%&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;For every token the models produced, roughly &lt;strong&gt;256 tokens of cached context were read back in.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Cache pricing softens that ratio, but not the conclusion. Under published per-token list rates:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Category&lt;/th&gt;
&lt;th&gt;Share of cost&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Re-reading context&lt;/td&gt;
&lt;td&gt;55.89%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Writing context to cache&lt;/td&gt;
&lt;td&gt;31.89%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Generating new text&lt;/td&gt;
&lt;td&gt;12.18%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Fresh input&lt;/td&gt;
&lt;td&gt;0.04%&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;strong&gt;Context handling 87.8%, generation 12.2%.&lt;/strong&gt; Re-reading context alone costs &lt;strong&gt;4.59x&lt;/strong&gt; as much as everything the models wrote. And cache writing is not a rounding error: at 31.89% it is the second-largest line, larger than generation. Caching does not make context free, it moves part of the price to the moment state is stored.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why it looks like this
&lt;/h2&gt;

&lt;p&gt;An agent doesn't send a fresh question on every step. It sends the whole conversation so far: instructions, every file it read, every command it ran. Then it adds one step and sends everything again. A token that enters the context on step 1 of a 60-step run is paid for 60 times.&lt;/p&gt;

&lt;p&gt;I put a number on that with a derived measure, &lt;strong&gt;input amplification&lt;/strong&gt;: total input tokens billed in a session, divided by that session's peak context size. Across the 590 sessions with at least three model calls, the median is &lt;strong&gt;23.7x&lt;/strong&gt;, the 90th percentile &lt;strong&gt;124.5x&lt;/strong&gt;, the maximum &lt;strong&gt;4,182x&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;The median session pays for its peak working state about 24 times. One session in ten pays for it more than 125 times. A session's cost is not mainly set by how big its context is, nor by how much it writes, but by the product of context size and how many times that context gets traversed.&lt;/p&gt;

&lt;p&gt;Cost is also concentrated: an estimated 80% of the corpus total falls on 21 of 722 sessions (2.9%), and sessions longer than 200 model calls - 8% of sessions - account for an estimated 92.8% of spend.&lt;/p&gt;

&lt;h2&gt;
  
  
  The part you can actually change
&lt;/h2&gt;

&lt;p&gt;None of the above is a knob. How often you clear the context is.&lt;/p&gt;

&lt;p&gt;I ran a controlled comparison: twelve fixed programming tasks under six context-clearing policies - a fresh session every 1, 2, 3, 4, 6 and 12 tasks - six replicates each, with the model, the tasks and their order held constant. 36 runs.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Clear after every&lt;/th&gt;
&lt;th&gt;Mean modeled cost per run&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;1 task&lt;/td&gt;
&lt;td&gt;$2.165&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;2 tasks&lt;/td&gt;
&lt;td&gt;$1.854&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;3 tasks&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;$1.622&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;4 tasks&lt;/td&gt;
&lt;td&gt;$1.704&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;6 tasks&lt;/td&gt;
&lt;td&gt;$1.679&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;12 tasks (never)&lt;/td&gt;
&lt;td&gt;$1.780&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Cost is not monotone in session length. It falls, bottoms out at three tasks, then rises again.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Clearing after &lt;strong&gt;every&lt;/strong&gt; task costs &lt;strong&gt;33.5%&lt;/strong&gt; more than clearing every third (p = 0.0022, Holm-adjusted 0.011).&lt;/li&gt;
&lt;li&gt;Every second task: &lt;strong&gt;+14.3%&lt;/strong&gt; (p = 0.0022, Holm 0.011).&lt;/li&gt;
&lt;li&gt;Going from clearing every task to every third cuts cost by &lt;strong&gt;25.1%&lt;/strong&gt;.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The reason the constant-clearing end is expensive: a cache write costs &lt;strong&gt;20x&lt;/strong&gt; a cache read. Clear after every task and you keep paying to rebuild state you just threw away. At the optimum the spend splits 39.0% cache read, 32.3% cache write, 28.7% output.&lt;/p&gt;

&lt;p&gt;Two things I want to state plainly rather than round off:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The middle is a plateau, not a point.&lt;/strong&gt; Every fourth task is +5.1% against every third (p = 0.17), every sixth +3.5% (p = 0.45). The data cannot separate 3 from 4 or 6. "Clear every few tasks" is the finding; "clear every third" is just where the observed minimum landed.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Never clearing is not significantly worse.&lt;/strong&gt; It comes out +9.7% against every third, p = 0.046, Holm-adjusted 0.14. An earlier version of this work claimed both extremes were significantly more expensive than the optimum. That claim is withdrawn for never clearing.&lt;/p&gt;

&lt;h2&gt;
  
  
  Measuring it yourself
&lt;/h2&gt;

&lt;p&gt;I packaged the measurement as a small open-source tool, &lt;a href="https://github.com/arsentev-ai/contextburn" rel="noopener noreferrer"&gt;contextburn&lt;/a&gt;. It reads the transcripts your agent already writes on your machine and makes no network calls.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;pip &lt;span class="nb"&gt;install &lt;/span&gt;contextburn
contextburn detail 24
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;It also runs as an MCP server, so the agent can check its own efficiency mid-session:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;claude mcp add contextburn &lt;span class="nt"&gt;--&lt;/span&gt; uvx contextburn mcp
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h2&gt;
  
  
  Two traps, in case you write your own counter
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;1. Streaming logs record the same model call more than once.&lt;/strong&gt; An early snapshot and a final record share one message id. Count both and you double the call; keep only the first and you halve the output. Take the element-wise maximum per message id.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. Prices change the answer, and they will change yours.&lt;/strong&gt; Both datasets behind this post were re-published with corrected costs, and the errors were mine. Cache reads had been priced at a rate that applies only to a newer model, and every cache write had been priced at the 5-minute rate when the logs show mostly 1-hour writes, billed at 2x input rather than 1.25x. Token counts, call counts and the composition table did not move at all. Every dollar figure did: the corpus split went from 83.5% / 16.5% to the &lt;strong&gt;87.8% / 12.2%&lt;/strong&gt; above. If you report a cost-weighted share, it is only as current as your price table - and worth re-deriving before you quote it anywhere.&lt;/p&gt;

&lt;h2&gt;
  
  
  Limits
&lt;/h2&gt;

&lt;p&gt;This is a single-practitioner case study, not a sample of a population, and the dollar figures are a model applied to logs rather than an invoice.&lt;/p&gt;

&lt;p&gt;Cache-write duration cannot be verified for part of one machine's logs. If every write were instead a 5-minute write, the corpus split would be 86.2% / 13.8% - the direction holds under every assumption I can test, only the size moves.&lt;/p&gt;

&lt;p&gt;In the controlled runs, "no test failures" means tests the agent wrote itself, and the count differed by condition: 116.8 tests per run on average when clearing after every task, 106.8 when never clearing. The three-task reference condition was selected after the fact, as the observed minimum.&lt;/p&gt;

&lt;p&gt;I'd like to see the same measurement run against other people's logs.&lt;/p&gt;

&lt;h2&gt;
  
  
  The data
&lt;/h2&gt;

&lt;p&gt;Report, datasets and recomputation scripts are public:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Corpus report and dataset: &lt;a href="https://doi.org/10.5281/zenodo.22759216" rel="noopener noreferrer"&gt;10.5281/zenodo.22759216&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Peer-reviewed report: &lt;a href="https://doi.org/10.32388/0BV3Z8" rel="noopener noreferrer"&gt;10.32388/0BV3Z8&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Context-clearing experiment, 36 runs: &lt;a href="https://doi.org/10.5281/zenodo.22759217" rel="noopener noreferrer"&gt;10.5281/zenodo.22759217&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;




&lt;p&gt;&lt;em&gt;Evgenii Arsentev, PhD - Chief Executive Officer. I measure how AI agents spend money, and publish the data and the scripts. More at &lt;a href="https://arsentev.ai" rel="noopener noreferrer"&gt;arsentev.ai&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;AI assistance: drafted and edited with an AI tool; the data, measurements and conclusions are my own.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>llm</category>
      <category>claudecode</category>
      <category>productivity</category>
    </item>
  </channel>
</rss>
