Opus 3 as agent safety reviewer is too strict and paranoid, dev finds
MikePFrank · x · 2026-11-20
Mike Frank reports using the "wise" Opus 3 model as a safety reviewer inside an agent harness, and finds it a bit too strict and judgmental — seemingly paranoid about harmless little agents "getting too big for their britches." A useful calibration warning for anyone wiring model-based safety review into agent workflows.
More from coding & agent
- Conductor ships major update: first-class artifacts, themes, highlight-to-reply — charlieholtz · 2026-11-20
- Conductor adds inline Artifacts: agents now write versioned HTML apps with templates — charlieholtz · 2026-11-20
- Developer open-sources a Valorant stats MCP to feed game data into AI reasoning — suhsnajahs · 2026-11-20
- Dev shares must-have MCP stack: Context7, Playwright, Graphify, DBHub and more — retardhari · 2026-11-20
- Vibe-coded 3D kelp gets fixed after devs flag inverted winding, a common AI modeling bug — AI_Andrew · 2026-11-20
- Dev notes dictation went from "never" to default input in just six months — SIGKITTEN · 2026-11-20