DEV Community: oVu X

Cut Your Coding Agent Bill by 18 to 31 Percent With This Microsoft Plugin

oVu X — Sun, 15 Nov 2026 15:15:22 +0000

What it is

TokenEconomizer is a free open source plugin. It cuts the cost of coding agents. A coding agent is software that writes and edits code for you. It works with Claude Code, Codex, the GitHub Copilot CLI, and Pi. Microsoft Research and UC Berkeley built it.

The code is on GitHub under the Apache-2.0 license. The project has 6,205 stars.

The plugin does not change your agent. It does not change the model. It learns from sessions your agent already ran. A session is one run of the agent in one repository. A repository is a folder that holds one code project and its history. The plugin uses that experience to cut waste in later sessions.

The paper reports clear numbers. On 277 SWE-Bench tasks from six repositories, mean cost fell 18.4 to 30.9 percent. SWE-Bench is a test set of real software tasks. The test used five agent and model combinations. Cost fell in all 30 repository and combination cells. Pass@1 rose in four of five combinations.

Pass@1 is the share of tasks the agent solves on the first try.

On Web-Bench, cost fell 16 to 21 percent. On two security-auditing benchmarks, cost fell 6 to 20 percent. Every task there sat in a new repository. Costs in these numbers include the plugin's own model calls.

How it works

Three mechanisms do the work. The first is micro-skills. A micro-skill is a short fact the plugin learned about a repository. Sessions keep rediscovering the same facts.

A build step must run first. A command has a spelling that works here. An environment variable is needed.

The plugin learns these facts from failures that repeat across sessions. A model also reviews recent sessions every three sessions and proposes facts. A fact is stated at the start of each session once three sessions confirm it. Later sessions can contradict a fact. Then the plugin drops it.

The second mechanism is tool pruning. Every loaded tool adds cost to every request. A tool that goes unused over 10 prompts across at least two sessions is left out. The same rule applies over five sessions of any length.

The plugin leaves the tool out through the agent's own settings. One command brings the tool back. Restored tools are never removed again.

The third mechanism is the loop guard. A turn is the agent's work on one prompt. A turn can get stuck in a loop. It repeats the same failures. The loop guard first learns how long a turn in this repository usually runs.

Then a judge model checks a turn that shows signs of looping. The judge must be at least 0.7 sure that the turn is stuck. Then the agent gets one redirect. A redirect is a short note that tells it to try another approach.

If the turn is still stuck, the plugin tells you how to stop it. It never stops a turn itself.

The plugin needs no API key of its own. Its model calls run through your own agent. They run on the same plan or login. Their cost is included in its reports. What it learns stays on your machine, in ~/.tokeneconomizer.

By default, tool pruning and the loop guard are on. Micro-skills is off by default. Its proposer adds model calls. You can turn each mechanism on or off per repository.

How to set it up

You need Python 3.10 or newer. You need one of the supported agents, signed in any way it supports. For Claude Code, run three commands. Clone the repository. Add the marketplace. Install the plugin.

Then start Claude Code and run /tokeneconomizer:status. The status command shows what the plugin learned and what your sessions cost. Other agents use the same commands with their own form. Codex uses $tokeneconomizer:status. Pi uses /te status.

Supported versions are narrow. Claude Code 2.1.300 to 2.1.310. Codex 0.153.1 to 0.164.0. Copilot CLI 1.0.83 to 1.0.102. Pi 1.2.0 to 1.3.0. The README says newer versions have not been tested yet.

Honest limits

Savings need time. The plugin starts with nothing learned. It learns only from sessions in the same repository. A repository you open once will not show savings. The gains in the paper came from repeated work in the same six repositories.

Some agents have partial support. Opencode, Cursor Agent, Grok Build, Hermes Agent, the Antigravity CLI, and Factory Droid are experimental. On the Antigravity CLI, only micro-skills without model calls works. Tool pruning is not available there. On Cursor Agent, session cost is unavailable. Check the README for your agent before you install.

The plugin's own model calls cost money. The fact proposer and the judge both call a model. Those costs are counted in its reports. That is why micro-skills is off by default. Turn it on only when you work in one repository often.

If any part of the plugin fails, your session carries on as if it were not installed. That is a safe design. But it also means a silent failure looks like no savings. Run the status command to check that it is active.

One more limit: the plugin works through what the agent already reads. It does not teach the agent new tools. It cannot fix a weak model or a bad prompt. It cuts waste. It does not add skill.

Verdict

TokenEconomizer solves a real problem. Coding agents repeat the same waste in every session. This plugin learns from that waste and cuts it. The measured savings are real: 18.4 to 30.9 percent lower mean cost, with success rates mostly up, not down.

The catch is patience. You need repeated sessions in one repository before the savings show. A developer who works in the same code base every day will gain the most. A developer who jumps between new repositories each day will gain little.

Try it if you use Claude Code, Codex, the Copilot CLI, or Pi. Install it in your main repository. Turn micro-skills on. Check the status command after a week. Watch the cost line before you judge it.

Follow along

I test AI tools and build in public:

GitHub: https://github.com/ovu-lab
YouTube: https://youtube.com/@oVuCyber

Your Team Chat Can Now Include AI Agents as Members

oVu X — Sat, 14 Nov 2026 15:14:21 +0000

What it is

agentes.chat is an open source team chat app. It works like Slack. The difference is that AI agents join as members.

They sit in channels and direct messages next to people. They read the channel. They answer when the work is theirs. They hand work back to people when it is not.

An agent is software that acts on a goal. Each agent has a role, an owner, and a monthly budget. The budget caps what the agent can spend on model calls. The owner gets the approval requests and the budget alerts.

The project appeared on GitHub on November 9, 2026. Five days later it had 1,375 stars and 241 forks. The repository lists 8,715 commits and 13 releases. That is fast growth for a project five days old.

How it works

Agents think with a model you choose. You can pick Claude, ChatGPT, Codex, Gemini, or Grok. You can also pick any model on OpenRouter. Your organization adds its own model key in the app. Model calls bill through your model key.

Agents connect to outside tools through MCP. MCP (Model Context Protocol) is the open standard that lets an agent call outside tools. The organization connects an MCP server once. Official servers for Linear, Sentry, Notion, and Stripe connect in one click. Each agent gets only the scopes you grant it.

Each agent gets its own computer and browser. The agent runs commands and code on that computer. The agent opens sites, fills in forms, and reads pages in its browser. You can watch the browser live in the chat. You can take over for a login or a captcha. Then you hand it back with a note.

A message with no @mention goes to the router. The router is the part that picks which agents answer. It can pick up to four agents, or none. The message says who picked. The router never grants a permission.

Merges, deploys, and anything a connector marks sensitive wait for a person. The approval shows as a card in the chat. The owner or an admin approves it. The approved action runs once, exactly as approved.

Agents keep decisions, facts, preferences, and lessons from conversations. They learn from corrections and refused approvals. They can propose changes to their own instructions. A person accepts or rejects each change.

An agent can also work on a schedule or on an event. For example, it can report every weekday at 9am. It can also act when a GitHub pull request opens. It delivers the result to a channel or a direct message.

The app also has threads, a wiki, a task board, and a notifications inbox. Agents can load skills from Markdown files. The whole app comes in 14 languages.

How to set it up

Self-hosting takes one command on a Linux server. Run curl -fsSL https://agentes.chat/install | sh. The installer asks first, then installs Docker if it is missing. It puts the app on your Tailscale network with private HTTPS. Tailscale is a tool that builds a private network between your devices.

The README says setup takes about 10 minutes. Postgres is the only dependency. Postgres is the database the app stores data in. Web, desktop, and phone apps connect to your server. The phone app works on iPhone and Android. The desktop app works on macOS, Windows, and Linux.

You can also use agentes.chat Cloud. Cloud is the paid hosted option. The README gives no price for Cloud. Self-hosting is free, under the AGPL-3.0 license. AGPL-3.0 is a free software license. There is no license key.

Honest limits

Several features are still in beta. The agent browser is beta. The Postgres connector is beta. Agent mailboxes are beta. The phone app is beta. The desktop app is beta.

Beta means the feature works but may still change.

The desktop app has a rough edge on first install. The Mac build has no Apple check yet. The Windows installer has no Microsoft signature yet. The first open asks once. The README says the download page shows each step.

You pay for the models. Self-hosting is free, but each model call bills through your own key. Each agent has a monthly budget. The budget is checked before every call. Set the budget low at first. Watch the spend before you raise it.

The project is five days old. That is very young. The code base is large and the team ships fast. Expect rough edges and fast change. Read the README and the issue list before you move a real team onto it.

The license matters if you change the code. If you change the code and do not publish your changes, you need the commercial license. The project sells that license for this case. Check COMMERCIAL.md in the repository if that applies to you.

Verdict

agentes.chat treats agents as team members, not as a sidebar bot. That is the right idea. Agents work where everyone can see them. Approvals and budgets sit in the open. The design puts the person in control of the risky steps.

The catch is age. The project is five days old. The core chat works. Several agent features are still beta. A small team that likes to test new tools can try it now. A team that needs a stable chat tool should wait and watch the releases.

Try self-hosting first. It is free, and it takes about 10 minutes. Set small budgets on each agent. Let the agents work in one channel before you give them the whole workspace.

Follow along

I test AI tools and build in public:

GitHub: https://github.com/ovu-lab
YouTube: https://youtube.com/@oVuCyber

Pinky Lets Your AI Agent Use Your Phone — With You in Control

oVu X — Fri, 13 Nov 2026 15:14:19 +0000

What it is

Pinky is a free, open source app. It lets an AI agent use your real iPhone or Android phone. You watch the agent work on your computer. You approve the risky steps. You can take over at any time.

Pinky appeared on GitHub on November 6, 2026. By November 13, it had 2,068 stars. That is fast growth for a project one week old. The project is pre-alpha. The author says so on the front page.

How it works

You install Pinky on your computer. Pinky runs on macOS, Windows, and Linux. You connect your phone with a cable or over Wi-Fi. You add Pinky to your agent with one click in Settings. Then you ask the agent to do a task on the phone.

For example, you can say: "Open Messages on my phone and tell Sam I am running late." The agent taps and types on the phone. You see each step in Pinky's window.

An agent here means software such as Claude Code, Codex, or Cursor. Pinky connects the agent to the phone through MCP. MCP (Model Context Protocol) is the open standard that lets an agent call outside tools.

The key idea is the approval gate. Payments, trades, and transfers wait for your click. You can make sends, orders, deletions, calls, installs, and settings changes wait too. These rules run in Pinky's gateway. The agent cannot approve itself. Screen text cannot talk the agent into it either.

Banking, wallet, and password manager apps stay hidden from agents by default. Agents cannot type into password or one-time code fields. Every new agent needs your approval first. Each agent gets only the scope you give it.

Pinky can also work without a second agent. Pinky on the phone does tasks by itself. You pick the model: DeepSeek, Qwen, Kimi, GLM, Claude, GPT, Gemini, or a local model in Ollama or LM Studio. On Android 13 and later, the agent works on a hidden second screen. Your own screen stays yours. It can stay off and locked.

Every run leaves a record. Pinky lists each step in plain words. It keeps screenshots and a screen recording. You can export the record as an HTML report. The record stays on your computer.

How to set it up

Android is the easy path. Turn on USB debugging in Developer options. Plug the phone in and tap Allow. Pinky offers to install its phone app. The app reads the screen in about 70 ms instead of 2 s.

iPhone setup takes more work. Plug the iPhone in and trust the computer. Turn on Developer Mode in Settings. Pinky mounts Apple's developer disk image on the phone. On iOS 27 and later, nothing gets installed on the phone.

On iOS 17.4 to 26, you install Pinky on the iPhone once from a Mac with Xcode. You sign it with your Apple ID. With a free Apple ID, the signature lasts 7 days. You install it again each week. iOS 17.3 and earlier do not work at all.

Honest limits

Pinky is pre-alpha. Some parts still wait for a person to try them. The author lists what works in docs/ACCEPTANCE.md. Read that file before you trust a task to it.

Wi-Fi is slower than a cable. The author measured a screenshot at 5 s over Wi-Fi and 1 s on a cable. A swipe shows after 80 to 210 ms over Wi-Fi and 60 to 90 ms on a cable. Short tasks may feel slow on Wi-Fi.

Screen sharing has side effects on iPhone. While Pinky streams the screen, some banking apps refuse to run. The camera shows black, so scanning a code fails. Pinky streams only while the phone page is open, while an agent works, or for 30 s after. Then those apps work again.

The Windows installer is not signed yet. Windows SmartScreen may warn you. You click More info, then Run anyway. That is normal for a young project. It is still a warning you should see with your own eyes.

Pinky costs nothing. The license is Apache-2.0. There is no account to create. Your model costs still apply. A call to Claude, GPT, or Gemini bills through your own account. A local model in Ollama or LM Studio costs nothing extra.

Verdict

Pinky solves the right problem. Most phone agents ask you to trust the agent. Pinky asks you to trust a gate instead. The gate sits outside the agent. The agent cannot change it. That is the honest way to give an agent your phone.

The project is still early. Android users get the smooth path today. iPhone users on iOS 17.4 to 26 face Xcode, signing, and a weekly reinstall. Wait for a later release if that sounds like work you do not want.

Try it if you like to test new tools. Start with the demo mode. It runs two scripted fake phones. You can see the approval flow before you connect your real phone.

Follow along

I test AI tools and build in public:

GitHub: https://github.com/ovu-lab
YouTube: https://youtube.com/@oVuCyber

United Airlines Just Banned AI Agents From Booking Flights. Here's What Actually Changed

oVu X — Thu, 12 Nov 2026 15:14:10 +0000

What happened

On Wednesday, November 11, United Airlines revised its Contract of Carriage to ban third-party AI agents from booking, changing, or servicing its tickets. It is the first major US airline to formally do so. The change was first reported by The Wall Street Journal.

The ban now sits in the contract's list of prohibited ticketing practices, alongside hidden-city and throwaway ticketing. A new Rule 6.J.6 prohibits "the use of AI Agents to make, modify or service a Ticket, and/or to purchase any service or product." A companion clause, Rule 5.H, gives United the right to cancel any reservation made, modified, or serviced by an AI agent — without notice.

The definition is deliberately broad. Under the new Rule 1, an "AI Agent" is any software or service — an AI assistant, a standalone agent, an AI-enabled browser — that takes autonomous or semi-autonomous action to discover, search for, book, purchase, ticket, service, or manage United products. And it applies whether or not you supervised or approved what the agent did. Watching your agent book the flight does not make it allowed.

The ban reaches beyond United's own website and app: United says it is restricting access "directly within our digital platforms and indirectly via travel agencies and aggregators, to prohibit use by unauthorized, third-party personal AI agents or assistants." That targets agents acting through those channels — booking a United fare yourself through an aggregator such as Kayak, Hopper, or Google Flights is not banned.

Why United says it is doing this

United's statement gives three reasons. First, handing your account login to a third-party assistant can expose sensitive customer information. Second, agent actions "may result in unintended purchases, account changes, or use of MileagePlus miles." Third, automated actions could conflict with FAA and Department of Transportation rules.

The check-in example makes the third point concrete. Under federal rules (49 CFR 175.25), airlines must tell passengers about dangerous goods they are prohibited from carrying, both when they buy a ticket and when they check in, and passengers must acknowledge they received that information. United's argument is that an agent completing check-in on your behalf could bypass that required acknowledgment.

What you can still do

The ban is about transactions, not research. An AI assistant can still search for and display United flights and prices — the purchase itself just has to be completed by the customer. AI-assisted trip research is still fine as long as the tool does not access United's website or app, and assistive technology for customers with disabilities is unaffected.

There is also a grace line for the past: United told Travel Weekly it will honor bookings made by AI agents before November 11, and will not cancel them on that basis. If you used an agent to book a United flight before the change, check that the reservation is correctly booked and ticketed, but the booking itself is safe.

The part United does not emphasize

The penalties are serious. According to an analysis of the revised contract by PPC Land, the remedies in Rule 6.K run from invalidating tickets and banning passengers to deleting miles, revoking Elite status, and ending MileagePlus membership. A cancelled reservation is the mild outcome; losing your miles and status is the sharp end.

There is also an obvious commercial side to this, and United's own leadership has not hidden it. CEO Scott Kirby said at a Chicago event that he did not want "to create another middleman." Aviation analyst Henry Harteveldt told CBS News that flight searches cost airlines money, and a surge of automated searches from personal agents could raise those costs without generating extra bookings.

That tension is worth naming plainly. The safety argument is real — the check-in acknowledgment is a genuine regulatory requirement — but the rule also protects United's control over its own distribution. This is, at least in part, an airline deciding that the checkout page belongs to it. That last reading is my interpretation, not United's stated reason.

Meanwhile, the industry is splitting rather than uniting. Delta was reported last month to be blocking similar tools, and Delta and American both told Fast Company they are still evaluating AI booking tools. Southwest went the other way last month, announcing a plugin partnership with ChatGPT to offer "a simpler, more direct path from inspiration and discovery to booking options." And while airlines restrict outside agents, they keep investing heavily in AI for their own ticket sales — enough that the Federal Trade Commission recently ordered Delta, United, and JetBlue to explain how they set fares and whether personal data plays any role.

The verdict

For travelers, the rule is simple for now: do not let an AI agent touch a United booking. Let it search, compare, and plan — then click "buy" yourself. If you already booked through an agent before November 11, your reservation stands, but double-check it is ticketed correctly.

For agent builders, this is the shape of the wall going up around checkout. The durable path is an authorized integration, and United says it currently works with no third-party personal AI agents — though it leaves the door open to approved ones in the future. Building an agent that books flights means negotiating with the airline, not just automating its website.

The bigger question is whether the rest of the industry follows. If Delta and American copy United's language, the "AI travel agent" idea gets squeezed into a world where agents plan and humans pay. If they don't, United's ban becomes a competitive quirk — and the agents just route around it. Either way, the line is now drawn in a real contract: AI can look. It can't buy.

Sources: United Airlines revised Contract of Carriage (revised November 11, 2026); The Wall Street Journal (first report); Simple Flying; Fast Company; CBS News. Penalty details via PPC Land's analysis of the revised contract. Interpretations of United's commercial motives are the author's, as flagged above.

Follow along

I test AI tools and build in public:

GitHub: https://github.com/ovu-lab
YouTube: https://youtube.com/@oVuCyber

48 AI Agents Scanned 1,200 km2 of Jungle Lidar - and Flagged a Maya Pyramid Missing From the Register

oVu X — Wed, 11 Nov 2026 18:40:02 +0000

What happened

In October, software engineer Pietro Miotti was traveling in Peru, reading about Hiram Bingham "discovering" Machu Picchu in 1911, when he had a very 2025 thought: what would Bingham do with today's tools?

He asked several AI agents (Gemini, ChatGPT, Claude and Grok). They all gave the same answer: lidar. Satellite photos of a jungle ruin show only treetops, but lidar fires laser pulses through gaps in the canopy and maps the ground underneath.

Peru turned out to be poorly covered by public lidar, so he pivoted to the Maya region, where roughly 1,200 km2 of lidar strips are publicly available. Then he went to bed and let the agents work.

How 48 agents "did archaeology"

There was no specialized archaeology model and no custom classifier. The lidar strips were converted into images, and 48 general-purpose agents went through them one by one, following a precise visual guide: look for pyramids, platforms and plazas; filter out natural hills, sinkholes and roads.

By morning they had flagged multiple candidate sites. One hilltop stood out: straight edges and sharp corners on the edge of a cliff - and nature doesn't really do straight lines. Zoomed in, it showed a levelled plaza, an oval pyramid roughly 8-12 m tall, a raised courtyard, and a flat-bottomed hollow that may have been an aguada, a rainwater reservoir.

The agents then checked the raw point cloud to confirm the shapes were in the ground returns, not vegetation artifacts - the site sits under 10-17 m of forest.

The official reply - and the correction

The agents drafted a formal report, translated it into Spanish, and Miotti sent it to INAH, Mexico's National Institute of Anthropology and History. On October 21, INAH replied: no registered archaeological site existed at that location. From random thought to an official government reply in under 72 hours.

Then, on November 11, Miotti published an update that every hype headline about this story leaves out: the site had already been mapped in Estrada-Belli et al. (2025). The agents missed that paper during their literature review. So this was a rediscovery - a real site, absent from INAH's register, already present in the academic literature.

Credit to Miotti for correcting his own story at the top of the post. That update is exactly what the agents got wrong, and exactly what a careful human has to catch.

What this actually proves

The bottleneck in this kind of work just moved. Scanning is now cheap, parallel and fast - 48 agents can "look" at 1,200 km2 overnight. What is still slow, human and error-prone is verification: the literature check, the official registry, the ground truth. The one clear failure in this project was not the agents' eyes. It was their fact-checking.

Barriers, if you want to copy the pattern

You need public data. This worked because Maya-region lidar is openly available; the original Peru idea died on coverage. No dataset, no project.
Verification is the hard part. General agents are good at visual triage and weak at exhaustive literature search - budget your effort accordingly.
Official channels are a gate, not a formality. In Mexico, fieldwork at archaeological sites requires INAH authorization; the finding is now in INAH's hands, as it should be.
There is nothing to buy. No tool, no subscription, no course. This is a workflow pattern: open data + many agents with a narrow checklist + a human who verifies and reports responsibly.

The takeaway

Most "AI agents" stories are still about coding. This one is a template for agentic work in any field with open data: chunk the dataset, fan out agents with one narrow job each, spend most of your own effort on verification, and route the result through the proper authorities.

The agents did the looking. A human still had to do the knowing.

Source: Pietro Miotti, "My AI Agents Found an Unmapped Maya Site" (canonical link above). Speculative interpretations of the site's function are the author's, as flagged in his post.

Follow along

I test AI tools and build in public:

GitHub: https://github.com/ovu-lab
YouTube: https://youtube.com/@oVuCyber