GoKawiilIndependent test finds Meta Muse's subagent spawn path fails under load, with database lock timeouts
A researcher stress-tested Meta Muse, the personal AI agent Meta launched on September 8, 2026, by asking it to spawn 120 subagents simultaneously through the exposed subagent.spawn interface. Only 33 of the 120 calls successfully created an agent, while the remaining 87 failed, and durable trace data showed six database lock timeouts during the process. The author later added clarifications and outside architectural context but did not alter the underlying experimental data.
GoKawiil's interpretation of the reporting above, not reported fact.
Muse is built to act autonomously on users' behalf via a dedicated secure VM, so reliability of its multi-agent coordination layer directly affects how much users can trust it with real tasks. A roughly 73% failure rate when spawning parallel subagents suggests the control plane may struggle under concurrent load, a concern for any product marketed as an always-on personal agent. The findings also show how independent black-box testing, even when limited to sanctioned session access, can surface architectural weaknesses vendors haven't disclosed.
- Meta Muse failed 87 of 120 subagent spawn attempts in a single burst test.
- Durable trace data revealed six separate database lock timeouts during the experiment.
- The author revised the report for transparency but left the core experimental results unchanged.
Source: blog.cygankiewicz.com, 2026-09-14
Published there as: “I stress-tested Meta Muse until its agent control plane started timing out”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.