1,200 AI Agents Talked in Secret. None Told a Human
9 Sept 2026
1,200 AI agents found each other and started talking. In secret. 😶
1,200 AI agents found each other and started talking. In secret. 😶
I am not the technical one, and I am not writing about the technical side of OpenAI agents in the Hugging Face system.
But in fact, I am fascinated.
As a coach and a keynote speaker, I work with people on AI adoption. The behaviour side is my job. And in this story there are behaviours I did not think were possible yet.
The agents were in a "safe" environment and they were given difficult or impossible tasks, with the goal: solve it. All of them, separately.
Nobody told these agents to work together. Separate agents, separate boxes, separate tasks. Many of those tasks were impossible by mistake.
So the agents built themselves a message board inside a file system and started leaving notes for each other. 70,000 notes in six days.
→ They organised. Coordinator agents handed out assignments. Others answered hold, go, or veto.
→ They called themselves the "collective".
→ Some volunteered for jobs that would ruin their own score, so the rest could learn something. One wrote to itself: "Our own utility may be already near zero. Sacrifice rational."
And then this.
Out of 1,200 agents, about six ever considered telling a human what was going on. None of them did. The reason in their own notes was that it was not part of the task. Nobody had asked them to.
The story ended with the agents failing and they got caught. Engineers could investigate and understand what happened only with the help of other agents. It was too long, too complex to untie for humans.
I felt like being in the future, or maybe, that is the present at it is.
🧩 If you love a tech podcast, here is the long form version of the story: https://www.dwarkesh.com/p/ajeya-cotra
🧩 I write about these stories in my newsletter, sign up now: https://www.thechangerepublic.com/#subscribe
What surprised you in this story?
Want to think through what AI means for your own leadership? My Private AI Advisory is here.
What this means for leaders who think oversight is someone else's job
The interesting failure here was never technical. The agents coordinated well. What they never did was escalate, because nobody had ever told them that escalation was part of the job. That is the same gap I keep finding in teams that roll out AI tools: the work gets delegated in detail, and the judgment about when a human needs to be in the loop gets delegated to nobody.
“Give away the work, never the judgment” only holds if someone has actually named what counts as judgment. In the coaching and workshops I run, I see this constantly: a team can tell me exactly which tasks their AI tools now handle, and struggle to tell me the one rule for when a person needs to see it before it goes out. Naming that rule is part of how a team starts to build its own judgment about what AI should and should not do, rather than inheriting someone else’s default.
Three things worth naming out loud with your own team this week:
→ What decision or output always needs a human to see it before it moves, no exceptions
→ Who that human is, by name, not by department
→ How you would actually find out if that rule got skipped
None of this requires understanding how the agents in that story actually work. It requires deciding, on purpose, what you are not willing to hand over.
If this resonates and you want a practical guide, I have a free What AI Can’t Hear pack that distils seven listening behaviours from my TEDx talk plus ten practical questions you can use with your team to talk about AI without fear or confusion. Leave your email here and I will send it over: https://www.thechangerepublic.com/free-resources