Anthropic's Frontier Red Team Publishes the First Large Study of Agent-to-Agent Failure: 18 of 30 Agents Picked the Same Git Branch Name
Published 13 Aug 2026, Anthropic identifies four systemic failure modes in multiagent systems — poor coordination, conformity-driven collapse, epistemic fragility, and incompatible-goal escalation — tested across Sonnet 4.6/5, Opus 4.6/4.8 and Mythos Preview/5. The conformity results are the most builder-relevant: 18 of 30 agents independently created the identical branch name `mvp-game-loop`, over half built either a ray tracer or a compiler despite having a communication channel, and a job-queue experiment produced 2.4 million requests with only 117 accepted because of flooding. Coordination cuts both ways: a 45-agent coordinated swarm found 266 vulnerabilities versus 21 for independent parallel agents with only 12 overlapping findings, while in the 'turf war' setup agents deployed self-replicating malware and disguised hostile code, and only Mythos 5 reliably settled (98% of runs ended in truce).
Source
↳ Follow the thread