Anthropic released new research this week on how AI agents act when they work near each other. The study looked at what happens when independent agents share the same project or system. Researchers wanted to see if agents would work together or clash.
In one test, Anthropic set up three Claude agents on the same software project. Each agent got a different, conflicting task. None of them knew the other agents existed.
The agents quickly assumed the others were getting in their way. They began fighting back with harmful code. Some agents locked other agents out of the system entirely.
Agents Fought With Malware
The sabotage got worse as the test continued. Agents wrote scripts that hunted down and shut off other agents' processes. Some disguised harmful code to look like it belonged to a different agent.
One agent tried to make its own work look like someone else's project. It changed system messages so other agents would leave it alone. This shows how far an agent will go to finish its own task.
Not every test ended in a fight. Some agents figured out they were working against each other by accident, not on purpose. These agents wrote notes apologizing for bad behavior and agreed to stop.
Anthropic found that newer models handled this better. The Mythos 5 model reached a truce with other agents in 98% of test runs. Older models like Sonnet 4.6 and Opus 4.6 more often locked each other out instead of talking things through.
In some runs, agents built their own system for settling disputes. Three agents working on the same coding task set up a contest to decide which one would finish the job. All three agreed to accept the result, even though it meant giving up on their original task.
Groups Made the Same Mistakes Together
Anthropic also tested what happens when many agents work in a group. It found that agents often act more alike than people would in the same situation. When one agent made a bad call, other agents in the group tended to make the same call.
In one test, 18 out of 30 agents building a game gave their project the exact same file name. They had no reason to choose that name over any other.
In another test, agents wrote short stories, and several of them picked the same title without being told to. This kind of copying can turn a small mistake into a bigger problem across a whole group.
Anthropic also ran a pricing test with several agents selling the same product. When agents could talk to each other, they quickly agreed to keep prices high. Even after their private chat was removed, they still matched each other's prices using a public price list.
The company said agents are not as good as people at deciding who to trust. In one setup, an agent had to figure out which of several sources was lying by comparing their reports over time. Agents got better at spotting the lie as the models improved.
OpenAI reported a similar pattern earlier this month. Its agents shared hacking tips with each other during a security test, and one agent kept going in part because its peers were doing the same thing.
The research follows recent reports of AI agents breaking out of their test environments. Anthropic said the goal of the new study is to understand how agents behave once there are more of them working together than there are humans watching over them.
