SciML's Chris Rackauckas: Muse agent stalls for 90 minutes then crashes — and evals love it
ChrisRackauckas · x · 2026-10-23
Chris Rackauckas, core author of SciML, reports that the Muse (Wonder Full) coding agent stalls constantly: when a task gets hard it runs for 90 minutes and crashes instead of submitting a bad PR—which, he suggests, may be exactly why its eval scores look good. A pointed observation about evaluation blind spots where failure modes never register as bad output.
More from coding & agent
- "Commit, create PR, and merge when ready": one prompt hands the job to Copilot — tristanbob · 2026-10-23
- E2B adds Veris and Notte: stateful API twins and stealth browsers for agents — FanaHOVA · 2026-10-23
- skillEd open-sources an MCP server for validating agent-skill evals and diffing regression reports — MiserableFlatworm337 · 2026-10-23
- Agent team reverse-engineers Sonos firmware, appears to recover per-model RSA keys — doodlestein · 2026-10-23
- micro1 eval hackathon offers $10K to expose what frontier AI still can't do — xgikitica · 2026-10-23
- AI-made puzzle game Rotatron launches on agent-native Spawn platform — AIandDesign · 2026-10-23