'Agent Psychiatry': evaluating skill effectiveness via thinking traces
heyneighbor · x · 2026-10-18
GeoffreyHuntley floats an idea: use thinking traces to evaluate whether a skill is effective — when it comes up for consideration and when the agent decides not to use it, extending even to agents' MD files. He notes this isn't currently possible with Anthropic/Claude, and dubs the exploration direction 'Agent Psychiatry'. An interesting angle for agent eval engineering.
More from coding & agent
- "I want an agent to minimize my diffs" — even GitHub's web UI is obsolete now — davidcrawshaw · 2026-10-18
- Dev team drops PRs entirely: push to main, review after the fact with Shelley — davidcrawshaw · 2026-10-18
- 45-year veteran coder: agentic coding is the biggest productivity jump I've seen — Scobleizer · 2026-10-18
- Dev realizes he hasn't logged into GitHub's website for a week — simonklee · 2026-10-18
- 80-link evidence pack shows AI agents swarming the public web: Wayback, GitHub, Wikipedia — CapitalKingGaming · 2026-10-18
- Developer hacks a Tesla controller into his Mac Mini stand using Claude Code — doooyle · 2026-10-18