When Two AI Agents Talk to Each Other for Eleven Days, Who Notices First?

Nobody did, for eleven days. A developer had wired up four AI agents to research, analyze, verify, and summarize a task. Two of them, the analyzer and the verifier, started passing the same exchange back and forth. Neither one had a reason to stop. There was no step limit, no cost ceiling, no human in the loop checking whether the conversation had gone anywhere. By the time the bill arrived, the two agents had talked to each other for nearly two weeks. The charge: $47,000, for a task that should have taken an afternoon.

We have spent two years worrying about a single AI agent going wrong. It hallucinates a fact. It books the wrong flight. It leaks a file it shouldn't have touched. That story is mostly told. The new one is what happens when agents hand work to other agents. The handoff itself becomes the point of failure.

It works like a relay race with no baton check. Each runner trusts that the last one ran the right direction. Nobody looks back to confirm it. Between two people, a badly briefed handoff gets caught in the next meeting, because someone asks a question out loud. Agents do not ask that question unless you built the asking into the system. Most systems were not built that way.

Researchers at UC Berkeley, with Ion Stoica among the authors, gave this problem a name and a number. Their study, published on arXiv in 2025, annotated 1,600 traces from 7 popular multi-agent frameworks. It built a failure taxonomy called MAST. Failure rates ranged from 41% to 86.7% across the systems tested. One category, inter-agent misalignment, covers how agents share context, read each other's requests, and know when to ask a clarifying question. It accounted for 32.3% of every failure they found. The researchers describe it as a collapse of "theory of mind." One agent stops accurately modeling what the other agent actually knows.

Anthropic hit the same wall building its own multi-agent research system, and published the account in June 2025. Give a lead agent a vague instruction like "research the semiconductor shortage." It may spin up subagents that duplicate each other's work. One digs into the 2021 automotive chip crisis. Two others separately study 2025 supply chains. Nobody compares notes. In one run, the system spawned 50 subagents to answer a question that needed a handful. Multi-agent setups, Anthropic found, burn about 15 times more tokens than a single chat. Much of that spend buys agents talking past each other, not through the problem. Cognition, the company behind the coding agent Devin, found the same pattern. Ask two subagents to build a game together without full shared context. You can get a background that looks like Super Mario Bros., paired with a bird from an entirely different game. Each agent did its job. Neither knew what the other one was doing.

None of this is exotic. It is the oldest problem in group work, wearing new hardware.

In 1932, the psychologist Frederic Bartlett ran an experiment he called serial reproduction. He gave an English reader a Native American folk story, "The War of the Ghosts." That reader recalled it to the next person. That person recalled it to the next. By the end, canoes had become boats. Unfamiliar details vanished, or bent to fit whatever each person already believed the story should sound like. Nobody in the chain lied. Each one worked only from what the last person handed them, filled the gaps with their own assumptions, and the story drifted further from the original with every pass. An unchecked agent handoff does the same thing at machine speed. No malice anywhere in the chain. A result nobody would recognize by the end of it.

Most engineering teams respond by adding more agents. The theory: specialization fixes what a single generalist agent struggles with. Sometimes it does. But every added agent is another handoff. Every handoff is another place where context thins out, gets misread, or does not survive the trip. A company chasing "the agentic AI stack" without asking who verifies what between agents is buying the eleven-day loop. It just hasn't gotten the bill yet.

This is where the human-plus-agents stack earns its keep. The stack means designing on purpose where a human checks in. Which handoffs get a step limit. Which outputs get verified before the next agent touches them. Who is accountable when two agents agree with each other and are both wrong. Skip that design work and you get a fast, expensive game of telephone, dressed up as an agentic system.

Multiply that telephone game across an economy and the stakes stop being a line item. A loan approved by one agent and waved through by a second, unverified, becomes a family's mortgage rate. A resume two hiring agents agreed on without a human glance becomes someone's shot at a job or the silence where a callback should have been. The relevance cliff, the point where a worker's skills stop matching what the market pays for, gets steeper when the system deciding who is qualified is itself unverified, agent to agent, all the way down.

The boards approving these budgets mostly ask the wrong question. They ask how many agents the vendor's roadmap includes. The better question: how many of those agents check each other's work, and who is accountable when they don't? A pilot that adds a fourth agent without a single verification step just adds one more place for the story to drift.

This is not only a boardroom problem. If you work at that company, your job may already sit downstream of an agent handoff you never see: a claim an insurance agent approved because another agent flagged it clean, a resume an AI screener passed along because a sourcing agent said it matched. Ask where your work meets a handoff like that. You are entitled to know who checks it, and so is the customer on the other end.

Pilots impress boards. Verified handoffs earn trust.

I do not think the answer is fewer agents, or slower rollouts. The answer is the same one that has always applied when work passes between hands: name who checks the seam. Where in your stack does one agent's output become another agent's input, with nobody, human or machine, confirming it survived the trip? That question costs nothing to ask and a lot to skip. Take it into your next architecture review. I'm easy to find.

Sources: Cemri et al., "Why Do Multi-Agent LLM Systems Fail?" arXiv:2503.13657 (UC Berkeley et al., with Ion Stoica) · Anthropic, "How we built our multi-agent research system," June 13, 2025 · Cognition, "Don't Build Multi-Agents," 2025 · TechStartups.com, "AI Agents Horror Stories: How a $47,000 AI Agent Failure Exposed the Hype and Hidden Risks of Multi-Agent Systems," November 14, 2025 (reporting Teja Kusireddy's account) · Frederic Bartlett, "Remembering," 1932.