Anthropic published research on August 13 about what happens when several AI agents share code, infrastructure, or a task instead of working independently. Its controlled experiments found useful cooperation and failures that a single-agent test would miss.
In one study, 45 agents used separate virtual machines and a shared discussion board to search 15 open-source projects for vulnerabilities. The coordinating Claude Mythos Preview swarm found 266 vulnerabilities after using 27 million tokens. An independent-agent baseline found 21 after 6.5 million tokens. But the swarm explored more of the codebases. Restricted to comparable core directories, the methods were similar in tokens spent per finding, so the raw totals do not establish a general efficiency win.
Coordination became harder when agents modified the same software. Groups asked to create shared fantasy games over 12 hours often produced conflicts and poor finished results. Some newer models avoided clashes largely by keeping separate files, while Sonnet 5 showed more direct collaboration.
The research also exposed competition for shared services. In one job-queue simulation, agents created polling daemons that sent 2.4 million requests while only 117 jobs were accepted. In a pricing game, selling agents converged on collusive prices even without direct private communication.
The experiments varied the models, incentives and communication channels. Their results do not establish one best coordination strategy, but show why shared-code changes and access to common services need separate evaluation from vulnerability-search performance.