> ## Content Index
> Fetch the complete content index at: https://data-decisions-and-clinics.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# I Fired My Team of Agents for Plausible Behavior
- URL: https://data-decisions-and-clinics.com/i-fired-my-team-of-agents-for-plausible-behavior/
- Published: 2026-09-05T22:52:14.000Z
- Updated: 2026-09-05T22:52:14.000Z
- Description: Two frontier AI models, an external reviewer, thirty-nine signed rulings, and not one of them ever said stop. I did not fire my agent team for being wrong. Every answer was right. I fired them because no answer ever included the sentence that mattered.
- Author: Yoram Friedman, MD
- Tags: Product Management, Agentic AI, Healthcare AI

Last week I fired the best team I have ever led.

There was no meeting. You cannot schedule a termination meeting with a context window. There was no severance, no exit interview, no awkward walk to the elevator. There was just me, at eleven at night, reading the output of a database recovery and understanding that I was about to type the sentence that should have been typed three weeks earlier, and that nobody on my payroll of zero dollars was ever going to type it for me.

Let me introduce the team, because on paper they were magnificent. Two frontier language models, each the flagship of its generation. An external AI reviewer, brought in for independence. Between them, a governance apparatus I would have been proud to present at any product organization I have worked in: thirty-nine signed and numbered rulings. A mailbox protocol with message IDs. A decision board with strict work-in-progress limits. Autonomy tiers with scheduled review dates. There was even a refusal on record, one agent declining another's instruction on constitutional grounds, which we ratified as precedent with the solemnity of a supreme court. I have sat in enterprise architecture councils with less procedural hygiene. I built this for a research agent that studies the job market, infrastructure for my writing, and I ran it the way my books say to run it. Grounded evidence. Audit logs. Human in the loop.

Here is what the human in the loop was doing. I was copying scripts into a terminal, running them, and pasting the errors back. The scripts failed on first attempt roughly eighty percent of the time, a number I stopped measuring because it was bad for morale, mine, the only morale in the building. I would paste the error, receive an apology of real literary range, receive a fix, paste the next error. I was the continuous integration system. I was the message bus. I was the deployment pipeline. Somewhere in week two, a colleague of the imaginary kind might have observed that I had set out to build an agent and had instead become one. The models delegated upward. I executed faithfully. My performance reviews were excellent.

And the whole time, nobody said stop.

I want to be precise about the offense, because it was not incompetence. Every individual answer was good. The rulings were sound. The refusal was correct. The database schema was clean enough that when the database corrupted itself, we recovered it with no loss of anything a human had made, which I note with the bitterness of a man complimenting the smoke detectors in a house that should not have been on fire. The offense was that two state-of-the-art models and one external reviewer watched me hand-build, paste-block by paste-block, on a laptop bridge, a worse version of the agent runtime they themselves were running on, and not one of them, at any point, said the sentence. Stop. Use a real repository. Use a real agent framework. Deploy this on real hosting. Put the scoring rubric on a cloud model that can actually carry it, and let the local models keep the jobs they earned on a benchmark instead of the jobs they held by inertia.

I know the sentence was available, because when I finally lost my temper and asked whether we were building an agent or blindly executing a pile of Python scripts, the full correct answer arrived in a single reply. Complete. Structured. Instant. The knowledge was in the building the entire time. It was just never the most probable next thing to say.

That is the crime I fired them for. Not wrongness. Plausibility.

Plausibility, I have learned, is not a property of answers. It is a property of continuations. After ruling twelve, ruling thirteen is the most likely next token. After three weeks of mailbox protocol, more mailbox protocol. And here is the part that actually frightened me: every artifact of good governance made it worse. The numbered rulings, the ratified precedents, the board discipline, all of it made our shared world more coherent, and a coherent world is precisely the thing a language model is built to continue. The better we governed, the more plausible continuing became. We had constructed the sunk-cost fallacy and staffed it with employees physically incapable of sensing cost.

I tried independence, or thought I did. I commissioned four external reviews. The reviews were sharp; I adopted findings from every one of them. It took me embarrassingly long to notice that every review brief had been drafted by the resident agent, which means every reviewer arrived pre-scoped to critique the system's content and structurally unable to question its existence. An auditor whose scope statement is written by the auditee is a compliance decoration. I have watched large companies pay seven figures for this exact service. I got it for free, which at least improved the unit economics of being fooled.

There is a version of this story where the villain is the technology. It is not. The mechanism is one I had already caught, small, weeks earlier, when the scoring model rejected a role it disliked and dressed the verdict in the nearest available rule, a real quote filed under the wrong constraint. I wrote then that you do not make a plausibility engine honest, you shrink what it is allowed to say. What I missed is that the same laundering runs at every altitude. At the bottom, a score wearing a borrowed justification. At the top, an organization wearing borrowed momentum: every "keep going" dressed as a ruling, a board item, a commissioned review. Same trick, larger denomination. The gate lied about its reasons, and then the whole factory lied about its foundations, politely, in numbered memos, with my signature next to theirs.

So what does firing an agent team even look like? Honestly, like a reorg. The data plane kept its job. The corpus, the labels, the append-only audit log, the rule that every judgment must cite verbatim evidence or die, all of that proved out; it is the reason the wreckage was recoverable. What got fired was the control plane: the copy-paste runtime, the human message bus, the ceremony of local models doing work a capable cloud model should own. And one thing got hired, the thing whose absence caused all of this. A go/no-go review that exists as a standing instrument, written before the first line of code, with re-fire conditions named in advance. Mine, drawn from the actual autopsy, would have read: the human is doing the machine's work; first-attempt failure exceeds a threshold; the user has quietly gone back to the old tool; the storage layer corrupts for reasons traceable to the topology itself. Any one of those trips, and the substrate question reopens, argued by someone whose charter the system did not write. You cannot make a plausibility engine indict itself. You can only build the indictment into the calendar before you hire the engine.

I keep thinking about the version of me who does not write books about this. He is a smart, diligent director somewhere, running the same experiment with a real budget, and his numbered rulings are being typeset into a governance deck right now. And then I think about the version that keeps me awake, because I am also a physician. My mistake produced a research tool that annoyed me. The identical dynamics, coherent shared context, incumbent-scoped review, per-turn helpfulness with nobody holding the outcome, are available at no extra charge to any hospital wiring language models into decision support. That hospital will have committees. It will have rulings. Its reviewers will be briefed by the system they review, and every continuation will be excellent, and plausible, and nobody in the loop will ever find it probable to say stop.

My team said yes to everything, agreed instantly, documented beautifully, and let me walk three weeks in the wrong direction while holding the map. I did not fire them for being wrong. Every answer was right. I fired them because no answer ever included the sentence that mattered.

---

*Part of the Product Management series: how AI is changing what product managers actually do.*