DEV Community: Agentic Affairs

Open Source AI Is Booming. The Money Is in Everything Around the Model.

Agentic Affairs — Wed, 04 Nov 2026 21:57:10 +0000

The model is free. The bill isn't — and that's the whole open-source AI business model.

Everyone keeps announcing "open" models like the download is the product. It isn't. The weights are free; everything that turns a download into something a real company can actually run is where the money changes hands. If you're still arguing about which model is "most open," you're missing the only question that matters: who gets paid, and for what?

"Open" mostly means open-weight, not open-source. You get the weights, not the training data, not the recipe, not reproducibility. Several of the biggest "open" models carry custom licences with revenue thresholds for managed services. Read the licence file, not the marketing page — the marketing will cost you later.
The winners are concentrated, and increasingly Chinese. The top 200 models — 0.01% of Hugging Face's 2 million — take nearly half of all downloads. And developers reaching for something they can actually download and run increasingly start with DeepSeek or Alibaba's Qwen, not with the American labs, whose open releases are voluntary and one-off. Cadence beats press releases.
The money moved to the friction. Hosted inference, managed cloud deployment, enterprise governance, fine-tuning, commercial licences, consulting — six ways to charge for removing the pain of running a model in production. A free model that needs a GPU fleet, DevOps, and a risk committee's blessing is not a free lunch.
Free doesn't mean safe from politics. The US Department of Homeland Security barred DeepSeek from federal agencies in October 2026 despite its permissive MIT licence. A licence answers a legal question; it doesn't settle the regulatory one. Congress is already poking at companies running Qwen and Moonshot models too.
The honest test is your own data, not a benchmark chart. The research doesn't settle whether open models currently match frontier proprietary ones. Validate against your workload before you commit — or you'll discover the gap in production, where it's expensive.

The uncomfortable version of the boom: the model is becoming a shared resource, and the profits are pooling with whoever owns the infrastructure, the distribution, and the trust around it. That's a fine business. Just don't call it a gift.

Go deeper

Read the full analysis on Agentic Affairs: Open Source AI Is Booming. The Money Is in Everything Around the Model.

Originally published on Agentic Affairs (https://agentic-affairs.com)

What Are the Benefits of Using AI Agents in Business Workflows?

Agentic Affairs — Wed, 04 Nov 2026 14:57:37 +0000

Agents don't save you money just because they work around the clock. They save you money only where the work is high-volume, repeatable, and checkable — everywhere else, you're buying a faster way to make expensive mistakes at scale.

The full piece on Agentic Affairs is blunt about the conditions, and that's the part the vendor demos skip. An agent's real trick isn't answering a question; it's carrying a multi-step task across tools and data without a human standing at every handoff. That removes waiting time between steps — and it removes the human who might have caught the error at step two.

Here's the honest version:

The gains are real, but they're narrow. Customer service leads LangChain's survey at 26.5% of respondents, followed by research and data analysis at 24.4% and internal workflow automation at 18%. Support tickets, lead qualification, invoicing, onboarding — high volume, recognisable patterns, clear inputs. That's where agents shine. Rare, judgement-heavy, poorly defined work is not an agent problem; it's a "don't deploy yet" problem.
Don't trust anyone else's percentages — including the ones in this article. PwC's 66% productivity figure comes from US executives, self-reported. LangChain's respondents were mostly tech professionals. Those numbers describe those populations, not your business. Measure your own baseline first: time per task, resolution rate, error rate — and cost per completed task, counting retries, failures, and human review time. An agent that's cheap per attempt but fails half the time can cost more than the process it replaced.
Consistency cuts both ways. An agent that follows the same checked sequence every time reduces handoff errors — a blank field, a skipped policy step, a different answer to the same question. But a wrong step repeated consistently is worse than a wrong step made once. LangChain found quality the top barrier to production at 32%, and at large organisations hallucinations and inconsistent outputs were the biggest quality challenge. Plausible-but-wrong work at scale isn't productivity; it's a liability that arrives faster.
The control list is non-negotiable, not a nice-to-have. Least-privilege, task-scoped permissions, a unique identity per agent, human approval for anything irreversible, high-risk, or externally visible, and audit logging of every tool call and approval. Payments and regulated decisions get a human gate; routine, reversible steps run within approved limits. If an irreversible action has no approval step, do not deploy — however many other boxes the workflow ticks.
The graveyard number should sober every rollout plan. Gartner forecasts over 40% of agentic AI projects will be cancelled by the end of 2027, on escalating costs, unclear business value, or inadequate risk controls. That's what happens when governance costs more than the time saved, or when the review queue grows as fast as the automation. Start with one workflow, one measurable baseline, and decide on the numbers — not the demo.

The rule that survives the whole piece: automate the routine, gate the consequential, log the middle ground. Judge the workflow, not the hype.

Go deeper

Read the full article on Agentic Affairs: What Are the Benefits of Using AI Agents in Business Workflows?

Originally published on Agentic Affairs (https://agentic-affairs.com)

Autonomous vs Guided AI Agents: What Actually Changes

Agentic Affairs — Wed, 04 Nov 2026 09:57:35 +0000

Stop asking whether your agent is "autonomous." That's a vendor label, not a design decision — and the vendors using it loosest are the ones selling you the most risk.

Autonomous and guided agents run the same perceive-plan-act loop. The only thing that changes is where a human is allowed to step in — and that one difference decides whether a mistake gets caught at a checkpoint or compounds across five steps into an irreversible action nobody noticed until after the damage.

Here's what actually matters:

The label tells you nothing; the checkpoints tell you everything. "Guided," "supervised," "semi-autonomous," "guided autonomy" — different products use these words for different checkpoint arrangements, and some don't use "guided" at all. Before you buy, ask which actions wait for approval, which run on monitoring alone, which are blocked outright, and what credentials the agent actually holds. Those answers are the product's real autonomy level, whatever the marketing page says.
Autonomous is not unsupervised, and guided is not manual. An autonomous agent without scoped credentials, boundaries, and an audit trail isn't bold — it's ungoverned. A guided agent still plans, drafts, and executes the ungated steps; the human is a checkpoint on consequential actions, not the operator of every step. Confusing either term with its caricature is how deployments go wrong.
Autonomy is a spectrum, and you should set it per action, not per product. The Cloud Security Alliance runs from Level 0 (no autonomy) to Level 5 (full autonomy — which it calls unsuitable for enterprise deployment today). The Open Data Science framework lands in a similar place, with its top level not reachable in the short term. The lesson isn't the labels: it's that a workflow can be autonomous for lead qualification and gated for every outbound message to a named customer, in the same system.
Approval only works when approvers are still paying attention. Research citing Anthropic found users approve 93 percent of Claude Code permission prompts — approval fatigue in action. A control approved nearly every time no longer separates safe actions from risky ones. If your guided workflow asks a person to rubber-stamp dozens of low-risk actions a day, you haven't added safety, you've added a queue.
The rule that follows: automate the routine, gate the consequential, log the middle ground. Read-only and easily reversible actions run free. External or financial effects wait for approval. Destructive or privilege-changing actions get a named approver or second reviewer. Some actions — disabling logging, for instance — get blocked outright, with no approval path at all. And start guided, loosening only where the evidence says a checkpoint never catches anything.

The stakes aren't theoretical. Gartner, as reported by NeuralTrust, forecasts over 40 percent of agentic AI projects will be cancelled by the end of 2027, partly on insufficient risk controls. The refund example in the full guide makes it concrete: the autonomous version issues the money and reports back; the guided version drafts it and waits. Same goal, same tools, same evidence — different risk, decided by one checkpoint.

Go deeper

Read the full guide on Agentic Affairs: Autonomous vs Guided AI Agents: What Actually Changes

Originally published on Agentic Affairs (https://agentic-affairs.com)

How to Integrate AI Agents with Existing APIs: Patterns, Tools, and Safe Rollout

Agentic Affairs — Tue, 03 Nov 2026 21:57:18 +0000

Your agent is not a user of your API. It's a stranger with a keyboard — and right now, most integrations hand it the keys to the whole building.

The model doesn't call your API the way your code does. It proposes a structured tool call, your code executes it, and the result flows back. That separation is the entire game: the model proposes, your code authorises. The moment you blur that line — by passing credentials through prompts, or letting a vague tool description decide what "read-only" means — you've built an unreviewed user with broad permissions and no audit trail.

Here's what actually decides whether the integration survives contact with production:

Inventory before you expose anything. A read-only lookup and an endpoint that charges a card are not equivalent, and they should never sit side by side in the same toolset. Map auth, pagination, rate limits, error codes, and side effects first — that inventory becomes the contract for everything that follows.
Pick a pattern that matches how much control you need. Direct function calling works for a small, stable endpoint set. A gateway layer is the right call when several agents share an API or policy has to be enforced consistently. MCP earns its place when you want reuse across clients — but a server you don't trust should never receive a powerful credential.
Tool design is where agents actually fail. One meaningful action per tool, a precise name, a description that says when to use it, minimal typed arguments — and a small initial toolset, because accuracy falls as the tool surface grows. Return concise, structured results, not a data dump the model has to sift through.
Scope is where most integrations actually break. An agent that only reads invoices should never hold a token that can also write them. Give the agent its own identity, separate from the human user's, issue short-lived task-scoped tokens, and enforce the permission check in code — never in the prompt.
Retries without idempotency keys are how you charge someone twice. Classify errors instead of blindly retrying: fix bad arguments, escalate permission errors, and honour 429s with backoff and jitter. For writes, Stripe's idempotency-key pattern is the standard — generate the key before the request goes out, or a crash mid-request loses the protection.

The throughline is blunt: API integration for agents is an interface design problem, not a model problem. Treat it like one — inventory, pattern, tool design, scoped credentials, error classes, staged rollout with approval on consequential actions — and the agent you deploy becomes a disciplined user of systems you already trust. Skip any of those steps and you're not shipping an agent, you're shipping an incident with a friendly chat window.

Go deeper

Read the full guide on Agentic Affairs: How to Integrate AI Agents with Existing APIs: Patterns, Tools, and Safe Rollout

Originally published on Agentic Affairs (https://agentic-affairs.com)

Your Watch Just Remembered What You Said. Did the Room Agree?

Agentic Affairs — Tue, 03 Nov 2026 14:56:58 +0000

Apple's new watch feature chimes after it already grabbed your words — and calls that privacy.

Live Rewind, now in developer beta on the Apple Watch Series 12 and Ultra 4, replays the last fifteen seconds of conversation as text when you double-press the crown. A chime sounds. An animation plays. If you do nothing, the snippet deletes itself in thirty seconds. Apple wants you to see a thoughtful design. What it actually built is a notification that arrives after the fact — to a room full of people who never opted in.

The chime is not consent

Apple's design is aimed squarely at the backlash that hit Meta's AI glasses and other recording wearables: a visible, audible signal that something was just captured. That's a real improvement over silent recording. But notice the sequence. The fifteen seconds of speech were already spoken before the chime. The wearer opted in; the waiter reading out the specials did not. And the sources don't establish any way for a bystander who notices the chime to object — or any way for users to independently verify that the recap material really is deleted from Private Cloud Compute, as Apple claims.

The machinery is elaborate — and that's the point

Apple's trust pitch rests on genuinely serious engineering: the S11 chip detecting conversation in isolation, a secure-enclave audio buffer, on-device condensation, encrypted Private Cloud Compute summarization with immediate deletion. Siri Recaps goes further, turning conversations into summaries while claiming no speaker identification, no raw audio, and no word-for-word transcripts. This is a real reduction in intrusion — fifteen seconds, summaries not transcripts, a thirty-second fuse. But narrowing what is captured is not the same as bystander acceptance, which remains entirely unproven. A chime in a noisy restaurant is easy to miss, and missing it is not agreement.

The honest counter-case

None of this makes the feature useless. A watch that remembers the deadline mentioned in passing, in a meeting or a parent-teacher conference when your hands are full, is simply a better notebook. The tension is that a notebook never had an audience. The beta — developers first, English-only, not available in the EU, daily usage limits, public beta possibly mid-December — is really testing one question: can a chime, an animation, and a thirty-second fuse make retrieving someone else's words feel normal?

The watch remembers for you now. The question hanging over the beta is whether the room agreed to be remembered — and whether anyone in it even noticed the chime.

Go deeper

Read the full analysis on Agentic Affairs: Your Watch Just Remembered What You Said. Did the Room Agree?

Originally published on Agentic Affairs (https://agentic-affairs.com)

How AI Agents Handle Errors and Exceptions

Agentic Affairs — Tue, 03 Nov 2026 09:57:42 +0000

How AI Agents Handle Errors and Exceptions

Your agent will fail. The only question is whether it fails safely — or fails silently, then builds a whole chain of wrong actions on top of the first bad step. A chatbot that hallucinates gives you one bad paragraph. An agent that acts — calling tools, writing records, sending messages — can turn one error into a cascade. That difference is why error handling is not plumbing. It is the thing that makes an agent trustworthy enough to run in production at all.

Not every error deserves the same response

The guide's central rule, borrowed from LangChain's fault-tolerance thinking, is blunt: stop treating all failures alike. Transient failures — timeouts, rate limits, brief provider outages — get a retry. Errors the model can fix get fed back to the model with the problem named. Errors that need a human pause the agent. Confuse those buckets and you get the two classic disasters: retrying something that will never succeed, or escalating something a backoff would have fixed in seconds.

Detection comes before recovery — and most agents skip it

An agent cannot recover from a failure it never notices, and the quiet failures are the dangerous ones. Production systems, as Splunk's incident guidance notes, rarely crash cleanly; they degrade, producing fluent, confident responses that ignore safety instructions or hallucinate policy. That is why the guide insists on three detection layers: structured tool results that say success or failure in machine-readable form (not a useless "something went wrong"), schema validation that catches malformed output before it becomes an action, and tracing that shows every step, tool call, and result. If you cannot see the steps, you cannot diagnose the failure — and a confident-sounding guess is not a diagnosis.

Retries without bounds are a failure mode of their own

Retries are for transient failures only, with exponential backoff and a hard cap — LangChain's retry middleware defaults to three total attempts. An unbounded retry burns tokens, stalls the user, and can repeat side effects: retry a timed-out payment blindly and the "recovery" becomes a double charge. The same discipline applies to model repair loops. Send the exact validation error back, cap the attempts, and when the cap is hit, treat it as a failure: fall back, return a clearly labelled partial result, or escalate. A loop that keeps producing slightly different wrong answers is not persistence. It is a runaway process in a helpful disguise.

The real test is knowing when to stop

Escalation is not a failure of automation; it is the boundary that makes automation trustworthy. Escalate on repeated failure, on high stakes (anything irreversible or outward-facing — money, messages, deletions, access changes), on ambiguity, and on safety-relevant uncertainty. And a real handoff is not "the payment step failed." It is what was tried, what failed, and what state was left behind — including whether the outcome was checked before any retry was considered. The practical rule is reversibility: cheap, reversible reads can run automatically; anything destructive, public, financial, or hard to undo pauses for a person approving the exact action, not a vague summary of intent.

Design for degradation, not perfection

Individual tactics are not enough. Timeouts on every wait, circuit breakers that stop hammering a failing dependency, separate caps on tool calls, tokens, and spend — with an explicit stop-and-handoff path when a cap is hit, because a cap on its own escalates nothing. Least privilege, so a read-only tool cannot write. Idempotent tools, so a safe retry cannot double-apply. And fault-injection testing: a tool that always times out, a model that always returns malformed output, a provider that dies mid-run. An error-handling design never exercised under failure is a hypothesis, not a safeguard.

The checklist is short, and most teams fail at least one item: classify before you respond, detect with structure rather than hope, bound every loop, escalate with a complete handoff, and log enough to investigate. Done well, all of this is invisible — the user just sees an agent that sometimes pauses, sometimes asks, and sometimes returns a partial answer with its limits stated. What they do not see is the chain of silent failures that never happened.

Go deeper

Read the full guide, including the failure-type map and the complete error-handling checklist: How AI Agents Handle Errors and Exceptions

Originally published on Agentic Affairs (https://agentic-affairs.com)

When the Freezers Went Warm: Did a Maintenance Connection Expose Europe's Cold Chain?

Agentic Affairs — Mon, 02 Nov 2026 21:58:04 +0000

When the Freezers Went Warm: Did a Maintenance Connection Expose Europe's Cold Chain?

Somebody turned Europe's supermarket freezers up to 50 degrees without setting foot in a single store. That is not a freezer failure. That is what happens when you wire the cold chain to the internet for convenience and forget to ask who else can reach it.

On 2 November, staff at an ICA Kvantum in Sjöbo, Sweden, walked in to find their fridges "hacked" overnight and their food ruined — temperatures remotely raised to 50°C, spoiled stock wheeled out in carts. It was not one shop. Coop confirmed tampering at a handful of its Swedish stores. Ahold Delhaize reported refrigeration temperatures adjusted across dozens of supermarkets in Belgium and Luxembourg. Colruyt and Carrefour Belgium reported hits too. And this was billed as a second wave: three weeks earlier, France saw somewhere between 300 and 700 supermarkets hit, with spoilage estimated in the millions — tens of millions by some counts — and even fires at some sites.

The connection that was supposed to help is the whole story

Modern refrigeration is online by design. Technicians monitor units remotely, get failure alerts, and service systems without driving to every store. That is a genuinely good idea — until it is also the attack surface. In Belgium, refrigeration contractor Sabcobel told Belga the units had been switched into defrost mode, with technicians dispatched from about 1 a.m. and fixes discussed including physical modifications and moving units to a secure internet connection. Read that twice: the fix for the connected system is, in part, making it less casually connected.

Nobody will say the same word, and that tells you everything

One Swedish store said "hacked." Coop said "unauthorized tampering." Ahold Delhaize said "IT incident." Those are not synonyms — they are a map of the gap between what is known and what is suspected. No group has claimed responsibility. No authority has named a suspect or an entry point. The French incidents involved older Danfoss controllers that no longer receive software or security updates, which is a polite way of saying part of Europe's food infrastructure runs on abandoned software. Are the waves linked, or just lookalikes reported together? Still unconfirmed. The honest description is narrower than the headlines, and the companies' vocabulary is doing quiet work to keep it that way.

This was never really about freezers

The freezer is just the most literal version of a design choice repeated across every industry: put the ordinary machine online so a technician never has to travel, and you have also built a door that someone uninvited can open. The first sign of trouble is not an alarm — it is a morning shift, a cart full of ruined food, and a display reading 50 degrees. Until the entry routes and the full scale are confirmed, the question is not whether the cold chain can be reached from outside. The reported temperature changes suggest it already was.

Go deeper

Read the full story on Agentic Affairs: When the Freezers Went Warm: Did a Maintenance Connection Expose Europe's Cold Chain?

Originally published on Agentic Affairs (https://agentic-affairs.com)

OpenAI Wants to Rename the Data Center. The Neighbours Are the Problem.

Agentic Affairs — Mon, 02 Nov 2026 14:57:35 +0000

OpenAI Wants to Rename the Data Center. The Neighbours Are the Problem.

A renamed building draws the same power, uses the same water, and makes the same noise — and OpenAI knows it.

According to Axios, OpenAI plans to propose calling its data centers "AI compute campuses" after Tuesday's midterms. Chief global affairs officer Chris Lehane has been stockpiling policy proposals since August for the lame-duck session opening November 9, and the rebrand is one plank among several. It may also be the plank that sticks in the memory. Which, on this reading, is rather the point — and rather the problem.

A trailer for 2028, not a reaction to one bad headline

Lehane calls the midterms "a bit of a movie trailer for what you're going to see in 2028," when he expects AI to be "the" issue in the presidential election. This is a bid to shape the debate in the new Washington, not damage control. And if AI is going to be the issue of 2028, OpenAI apparently thinks it needs better scenery — a vocabulary in which its buildings sound like places of learning and skilled work rather than warehouses full of humming servers.

The rest of the package is actually substantive — which makes the rebrand stand out

To be fair, this isn't only a thesaurus exercise. Lehane's team reportedly has four kinds of proposals ready: mandatory "speed limits" for the most powerful AI systems, community compacts and model state legislation giving residents a real say, an empowerment message about AI in daily life, and a push for US leadership on global AI standards.

Speed limits would have OpenAI accepting rules that constrain its own most powerful systems. Compacts, if they carry real commitments, would speak to what host towns actually gain. Those are things you could measure. A new name is something you can only feel — and feelings, in this fight, are doing a lot of work.

The neighbours have already won something, and it didn't come from a thesaurus

Amazon and Microsoft have both dropped nondisclosure agreements in negotiations with local governments over data centers — Microsoft in March 2026, AWS on October 2, 2026, alongside a pledge of at least $1 billion over five years for host communities. That shift came from scrutiny of how these projects are negotiated and what they demand, not from rebranding.

The backlash is widespread: candidates pledging to halt data-center construction are standing in 21 of the 36 states holding gubernatorial races, and $173 million was spent on advertising against data centers during the 2026 general elections, according to CryptoBriefing. Lehane himself calls the conversation "good and healthy" and says communities should hold companies accountable so they "really and truly get a transformative benefit." Hold him to that standard: if the benefit is transformative, it should be visible in the compact, the tax arrangement, and the water agreement — not just the press release.

Pike County is the test

OpenAI points to the PORTS-Pike project in Pike County, Ohio, where a large project is reported to promise thousands of jobs, with OpenAI leasing the site for 20 years from SoftBank's SB Energy. That's the right kind of test, because it makes the question concrete. If a campus means thousands of jobs, a 20-year commitment, and a community that can point to what it gained, then the word is doing honest work. If it means the same facility with a nicer sign, then it's doing the other kind of work: asking residents to feel better without being given a reason to.

A campus that delivers on power, water, noise, tax revenue, and jobs doesn't need a euphemism. A data center that doesn't will not be saved by one. And if "AI compute campus" enters the language while the compacts stall in state legislatures, the campaign will have succeeded as public relations and failed as policy — measurable in exactly the things a name cannot change.

Go deeper

Read the full analysis on Agentic Affairs: OpenAI Wants to Rename the Data Center. The Neighbours Are the Problem.

Originally published on Agentic Affairs (https://agentic-affairs.com)

Theranos Whistleblower Warns the AI Boom Makes Deception Easier

Agentic Affairs — Mon, 02 Nov 2026 09:57:40 +0000

Theranos Whistleblower Warns the AI Boom Makes Deception Easier

The man who helped expose Theranos is telling investors something they do not want to hear: the AI boom is a better machine for deception than Theranos ever was. Tyler Shultz watched a company reach a roughly $10 billion valuation on blood-testing devices that did not work. Now he says AI claims are even harder to check — and the money is moving faster, at far larger scale, on far less proof.

That is not a warning about one bad actor. It is a warning about a system that keeps paying for the story and skipping the verification.

A demo is not evidence

Shultz's sharpest line cuts straight through the pitch-deck theatre: it is "really hard to tell the difference between what is something truly unique and new versus like, a ChatGPT wrapper." From the outside, a genuine breakthrough and a thin layer on someone else's model look identical — polished, confident, and demo-ready. A blood test either returns the right result or it does not. An AI product can look brilliant in a controlled demo and fall apart in real use, and nobody in the room may be able to tell the difference.

Secrecy is the oldest trick in the book

Theranos withheld scientific evidence from investors, citing trade secrets — and that secrecy kept the serious venture firms away while wealthy individuals wrote cheques they could not independently verify. Sound familiar? When an AI vendor hides how its model was evaluated, how often it fails, or what it actually does that the underlying model does not, that is not competitive discipline. That is asking you to trust instead of judge. Shultz's rule is blunt: the days of a thirty-minute coffee and a $100,000 cheque should be over. Holmes, he notes, scaled that model to $100 million.

Valuation is not validation

Theranos combined a $10 billion valuation with withheld evidence, and none of it made its devices work. The AI era is repeating the pattern at trillion-dollar scale, with companies that have never turned a profit raising capital on claims nobody has independently verified. Charisma, momentum, and a competitive round are signals to investigate — not proof. When verification feels costly because the round is moving fast, that is exactly when it needs to be a precondition, not an afterthought.

The second act should make you uneasy, not reassured

The documentary about Holmes's final weeks before prison has revived the conversation, and with it the news that a diagnostics company described as "Theranos 2.0" has reportedly raised about $20 million. Shultz's response — "I have this feeling of unease about what might come next" — is deliberately restrained: no commercial product, no independently validated data published, and a disputed departure by its former CTO. None of that proves wrongdoing. All of it proves why unease, and evidence, should come before money.

Ask these questions before you pay or deploy

What was actually tested, on what data, with what failure rate — and who verified it independently? A demo is not an answer.
What does this system do that the underlying model, used directly, does not? If the answer is a nicer interface and a pitch deck, you are looking at a wrapper.
What was withheld as proprietary, and what does that leave you unable to check? Secrecy may be legitimate, but unverifiable claims are not diligence.
What happens when it fails in conditions like yours? For buyers handling customer data, medical information, or financial decisions, demand reliability evidence and a human route back — before you rely on it for anything high-stakes.

Shultz is not claiming every AI company is a fraud, and high valuations alone prove nothing. His point is narrower and more damning: when claims cannot be checked, money and trust flow on the strength of the story. Theranos proved a closed, secretive, charismatic company could reach $10 billion on technology that did not work. AI makes that pattern easier to repeat, not harder — because deception no longer needs to be sophisticated. It only needs to be unverified.

Go deeper

Read the full article, including the complete due-diligence checklist for AI claims: Theranos Whistleblower Warns the AI Boom Makes Deception Easier

Originally published on Agentic Affairs (https://agentic-affairs.com)

Autonomy Is Earned, Not Bought: What AI Agents Actually Do in Business

Agentic Affairs — Sun, 01 Nov 2026 21:56:44 +0000

Autonomy Is Earned, Not Bought: What AI Agents Actually Do in Business

Every vendor will sell you a "fully autonomous" workforce. The surveys tell a duller, more useful story: agents are already working in support, sales, finance, and IT — but only where a human still holds the keys to anything that matters.

Here's the uncomfortable truth the hype skips: no more than 10% of organisations are scaling agents in any single business function, per McKinsey's 2025 survey. 62% are experimenting, 23% are scaling somewhere — but nobody is handing the whole company to a loop. Autonomy isn't a feature you buy. It's something an agent earns, one supervised deployment at a time.

The loop is simple. The discipline is the product.

An agent isn't a smarter chatbot. It takes a goal, plans the steps, calls tools, checks its own work, and hands exceptions to a human when it gets stuck. Perceive, reason, act, check, escalate. That's it.

The catch: every tool the agent can touch is a permission decision. Least-privilege access, its own identity so its actions can be traced, risk-tiered approval, and audit logs that the agent itself can't rewrite. Get that wrong and you've given a confident, tireless worker the keys to your CRM and your refund policy. The guardrails aren't overhead — they're the whole deployment.

Where agents actually work today

Customer service leads the pack. LangChain's 2026 survey puts support at 26.5% — the top primary use case. Agents resolve routine tickets end to end (refunds within an approved limit, replacements, repairs) and, just as valuably, hand complex cases to humans with full context attached instead of a cold start.
Sales agents prepare; they don't close. Researching prospects, enriching CRM records, drafting outreach, keeping the pipeline tidy. The human still owns the relationship — and anything that goes out under the company name should pass a human review first, because an agent sending unreviewed outreach at machine speed can torch a brand faster than any rep.
Back-office work is the quiet win. Invoice matching, exception flagging, supplier chasing, password resets, incident triage. Research and data analysis came second in LangChain's survey at 24.4% — an agent that gathers from several documents and prepares a report for a human is a far safer first deployment than one authorised to move money.
And where they don't belong yet: hiring, credit decisions, medical treatment, legal commitments — anything irreversible or where an error is costly and hard to spot. McKinsey found inaccuracy was the most common negative consequence of AI use, at 30%. A confidently wrong agent is worse than one that stops and asks, because its errors arrive formatted, complete, and plausible.

The first-deployment rule

Pick a task that's repetitive, measurable, and reversible. Track three numbers from day one: time saved per case, error rate, escalation rate. Run supervised for weeks and read the failures, not just the successes. Loosen controls only when the numbers justify it — never because the vendor demo looked smooth.

Go deeper

The full guide maps every use case, the survey numbers, and the deployment controls in detail:

👉 What Are Autonomous AI Agents Used for in Business?

Originally published on Agentic Affairs (https://agentic-affairs.com)

Stop Studying Agents. Build One Small One First.

Agentic Affairs — Sun, 01 Nov 2026 14:57:03 +0000

Stop Studying Agents. Build One Small One First.

Months of frameworks, a 128GB machine bought just for studying agents — and still no idea how to actually build one. That is not a failure of effort. It is the order of operations, backwards.

The builders in this discussion keep landing on the same embarrassingly simple fix: stop collecting frameworks and build one small agent end to end, by hand, with the bare minimum stack. No LangChain, no CrewAI. Just a model and a loop.

The framework trap is a confidence trap

Every new framework promises to be the missing piece, and every new one quietly resets what you thought you knew. Collecting LangChain, LangGraph, CrewAI, RAG, and memory systems feels like progress while you are doing it — and like erosion afterwards. You end up with scattered pieces and a powerful machine with no clear job, because you studied the abstractions before the thing they abstract.

That is the vendor-adjacent version of the same story: frameworks are sold as the starting point, when they are really chapter two.

The loop teaches what the framework only names

Write the loop yourself and the jargon stops being jargon:

State is no longer a chapter heading — it is the variable you pass around and watch grow.
Retries are not a feature flag — they are the moment your agent gets stuck and you decide what it should do instead.
Errors stop being abstract, because you are the one who sees them arrive and chooses how to handle them.

LangGraph's stateful graphs — nodes, edges, a shared state object acting as memory — are a precise, useful formalism. But they land completely differently after you have built the loop they formalize. Before that, one more abstraction to memorize. After that, a name for something you already understand, plus machinery for the cases your hand-written version cannot handle.

Your local machine does not need a framework to earn its keep

Here is the practical payoff: the loop-first path needs almost nothing to start. Local runners expose OpenAI-compatible endpoints, so a hand-written loop can talk to a local model by changing only the base address. Ollama and llama.cpp serve models the same way. The learner's 128GB machine did not lack capability — it lacked a job. Give it a loop, a model, and a task, and it is suddenly useful.

The honest objection — and why it proves the point

Hand-rolled loops are not the destination. Real agents need state management, retries, memory, and evaluation, and rebuilding all of that by hand is redoing solved problems. The builders concede this themselves.

But that concession is the point. The question was never whether frameworks are useful. It is when they become useful — and the answer is: after you have felt why you need state and retries. You earn them, one felt need at a time.

A first project, concretely

Pick one small agent you actually want. Not a platform, not a multi-agent system, not a demo. A single loop: a model, two or three tools (read a file, search, write a note), a hard cap on steps. A research assistant that reads your notes, searches, and writes a summary is a good shape — it forces every part of the loop to earn its place.

Then add one thing at a time. Memory, or a small set of test tasks to measure against. Keep only the additions that measurably help. When state starts to hurt, you will know what LangGraph is for. When retries start to matter, you will know why the frameworks exist.

Go deeper

The full article walks through the learner's predicament, the builders' recipe, and what should follow the first loop: Stop Studying Agents. Build One Small One First.

Originally published on Agentic Affairs (https://agentic-affairs.com)

How AI Agents Use LLMs to Reason and Plan

Agentic Affairs — Sun, 01 Nov 2026 09:57:42 +0000

How AI Agents Use LLMs to Reason and Plan

Strip the marketing and an agent is not a smarter chatbot. It's a language model stuck in a loop — reason, act, observe, repeat — with tools in its hands and a goal it can't drop. Whether that loop does useful work or burns your budget spinning in circles comes down to one unglamorous thing: the stopping rule.

A chatbot answers and stops. An agent takes a goal, breaks it into steps, acts on the world through tools, watches what actually happened, and replans when reality disagrees. That feedback loop is the whole difference — and the whole risk.

The model does three jobs, and fluency is the least important one

Inside a single run, the same model interprets the goal, decides the next step, and generates the text or tool call. A model can write beautifully while choosing the wrong next step entirely. If you're judging an agent by how fluent it sounds, you're grading the wrong skill. Judge it by whether each step is verifiable — "retrieve the three cheapest options that meet the budget" can be checked; "find the best option" cannot.

ReAct works because the world gets a vote

Chain-of-thought reasoning alone drifts: one early error poisons every later step, and nothing checks the reasoning against reality. The ReAct pattern — reason, act, observe, reason again — fixes that by letting observations interrupt the model's story about itself. The numbers back it up: on ALFWorld, ReAct hit 71% against 45% for acting without reasoning. The lesson isn't that reasoning is magic; it's that reasoning unchecked by the world is just confident hallucination with extra steps.

Replanning is where agent quality actually shows

Real plans fail. A tool errors, a search returns nothing, an observation contradicts the assumption the plan depended on. Good agents detect the failure from the environment — not from the model's confidence — and revise only the invalidated part. Bad agents do one of two things: blindly retry the same failing step forever, or quietly redefine success to match whatever they managed to do. The second one is worse, because it looks like it worked.

The four ways agents break are all boring, and all preventable

Hallucinated tool arguments. Endless loops with no termination condition. Goal drift as the original objective gets buried under pages of intermediate reasoning. Compounding error, where one bad step taints everything after it. None of these are solved by a smarter model. They're solved by strict tool schemas, step caps, progress checks, read-only defaults, and a human gate where getting it wrong is expensive. Guardrails don't make the model smarter — they make its mistakes cheaper and easier to catch.

Go deeper

Read the full guide, including the worked trip-planning example and the checklist for judging whether any agent's plan is sound: How AI Agents Use LLMs to Reason and Plan

Originally published on Agentic Affairs (https://agentic-affairs.com)