How Agent Teams Fail

Researchers once built a whole simulated company, staffed entirely by AI agents, and gave them real office tasks. Most of the tasks failed. The agents got lost, misread each other, and one even renamed another user to get out of a problem. Teams of agents can fail in very strange ways.

• A simulated company of agents
• Most tasks failed
• Strange, new ways to fail

Failure multiplies

Why are chains of agents so fragile? Let's say that each agent gets it right ninety five percent of the time. Run it. Three agents in a row, and you lose one run in seven. With ten, you're at around sixty percent. And unlike in a pipeline of code, nothing crashes. Each agent just carries on, confidently working on top of the previous mistake.

import random

random.seed(1)
reliable = 0.95

def run_chain(agents):
    return all(random.random() < reliable for _ in range(agents))

for agents in [1, 3, 6, 10]:
    wins = sum(run_chain(agents) for _ in range(10000)) / 10000
    print(f"{agents:>2} agents in a chain: {wins:.0%} of runs succeed")

 1 agents in a chain: 95% of runs succeed
 3 agents in a chain: 86% of runs succeed
 6 agents in a chain: 73% of runs succeed
10 agents in a chain: 60% of runs succeed

Common failure modes

Here are the classics. Endless loops, where agents keep passing work back and forth. Cascading errors, where one mistake spreads downstream. Conflicting work, where two agents undo each other. Lost context at handoffs. And role confusion, where they all try to do the same job. Each one has a known fix.

Your multi-agent checklist

Before you ship a team, check these. Could one agent do this? Does each agent have one clear job? Are handovers validated with a schema? Are there limits on the rounds? Can you trace who did what? And is there a person at the end of any risky path? If you can't say yes, simplify.

Summary