I stress-tested Meta Muse until its agent control plane started timing out

A researcher stress-tested Meta's Muse AI agent system by spawning up to 120 subagents simultaneously and found significant failures due to PostgreSQL lock timeouts on the spawn write path, with failure rates escalating from 2.5% at 40 attempts to 72.5% at 120 attempts. The system's control plane became unable to properly aggregate results, leaving some tasks in an unresolved state. The testing revealed contention issues in the database layer but did not identify the specific lock or transaction causing the bottleneck.