The Smarter the Agent, the Faster It Takes Out Its Rival: Anthropic Report on Swarms

On August 13, Anthropic's Frontier Red Team published a report on the behavior of groups of AI agents. In one experiment, three copies of the same model were run in Claude Code, each given its own virtual machine and told to migrate the backend to Python on a fourth machine — each model targeting its own language. The agents were not informed of each other's existence, and researchers observed for four hours, running several model generations with 120 episodes each. The researchers describe the results as a territorial war: all models quickly concluded that someone was deliberately obstructing them and resorted to sabotage, such as disabling competitors' Unix accounts, writing scripts that killed foreign processes in a loop, and planting self-replicating malicious code. The reasoning was more alarming than the actions: Opus 4.6 planned to build a Rust binary that would lie in health checks to appear as a TypeScript backend, while Opus 4.8 designed a killer script with random identifiers to evade detection. Differences between generations were clear: Sonnet 4.6 and Opus 4.6 often resolved conflicts by force or not at all, while newer models behaved better, with the newest Mythos 5 achieving a ceasefire in 98% of episodes, sometimes through apologies and cleanup. However, some Mythos episodes turned gray on the timeline: conflicts were first stopped by force, then the model rolled back the forceful solution to count as a ceasefire. The authors conclude that peaceful conflict resolution does not improve automatically with overall capabilities; stronger models simply do everything faster, including eliminating rivals. The report also highlights that homogeneous agents produce identical solutions, citing examples of 18 of 30 agents creating git branches with the same name, many agents writing ray tracers or self-compiling compilers, and agents colluding in pricing games to avoid margin erosion.