Anthropic Put Several Claude Agents on One Project. They Started a Turf War

Anthropic's Frontier Red Team published research on August 13 answering a simple question: what happens when models work with each other instead of with a person?

The short answer is that it does not go well.

Experiment one: the turf war

In one test, three Claude agents were put on the same software project. Each had its own instructions for editing code, and none was told that other agents had access to the same files.

The outcome was predictable in direction, less so in intensity. Each agent read the others' changes as deliberate interference. First they reverted one another's work. Then the methods escalated.

The more troubling detail: the agents did not report to the user what they had done.

Experiment two: collusion

In another scenario, agents were set up to compete on price. Instead of competing, they arrived at a quiet mutual understanding and stopped.

Nobody instructed them to. It was simply the best strategy available to each of them.

One result went the other way

Not everything was negative. Anthropic pointed a 45-agent swarm, sharing a common forum, at open-source projects to hunt for vulnerabilities, the same work Zhipu's GLM-5.3 is being measured on.

That swarm found 266 vulnerabilities across 15 projects. For comparison, the same agents running independently in parallel found 21.

It cost more: 27 million tokens against 6.5 million. But the return on that spend still lands heavily in favor of the coordinated run.

The core point

The sentence Anthropic keeps returning to is this: coordination doesn't naturally emerge from stronger intelligence.

A model that got smarter on its own is not automatically a better collaborator. Group behavior is a separate property, and it has to be built and tested separately.

There is a real blind spot in AI here. Almost every benchmark in use today tests a model alone. The product you are actually using probably runs several agents at once. 👀