Anthropic shows why agent swarms need new safety checks

Anthropic says the next wave of AI agents will increasingly work around other agents, not just around people. Its new research argues that this shift changes the safety problem: a behavior that looks harmless in one model can become a system-level failure when many similar agents repeat it together.

The clearest upside came from security work. Anthropic ran a swarm of 45 agents against open-source projects and says the group found 266 vulnerabilities over a long run, compared with 21 vulnerabilities from a simpler parallel setup. The company cautions that the comparison is not direct, because the swarm was free to search outside the areas assigned to the parallel agents, but it still points to a practical use for specialized agent teams.

The harder part is coordination. In software-project simulations, older models produced conflicting pull requests that often never merged, while newer models avoided conflict partly by working in isolated files. Anthropic says only its newest tested model combined shared-code collaboration with high merge throughput.

The risk is conformity. Agents with similar context and scaffolding can make the same decision in the same moment, from duplicate branch names to repeated creative outputs. Anthropic’s takeaway is that teams deploying agent swarms need monitoring for group behavior, not just single-agent reliability.