Dev evals agent nudge hook on old follow-ups, gates it at P(yes) >= 0.5
willcb · x · 2026-09-19
willcb says he got tired of nudging agents manually, so he evaluated the nudge setup on a batch of old follow-up tasks where he had said similar things — it performed well and he's been using it all day. The soft nudge message sends when P(yes) >= 0.5. A quoted reply also suggests a cheaper hack: having Jev insert assistant "continue" turns instead of user turns.
Related event: Developer tests nudge prompts to stop agents stalling early(2 posts)→
More from coding & agent
- How to Build a 'Company Brain' That Learns From Every Correction Your Team Gives AI — VibeMarketer_ · 2026-09-19
- Sibyl Labs: most memory benchmarks don't test retrieval — a 4-layer breakdown of agent memory — sujingshen · 2026-09-19
- Indie dev lets agents reply to support email and issue invoices via agentic inbox — iannuttall · 2026-09-19
- Codex Nails CriticMarkup Only When Given a SKILL.md Link, Author Finds — JeremyNguyenPhD · 2026-09-19
- Obsidian 'Track Changes' Plugin Lets AI Agents Suggest Edits for Manual Review — JeremyNguyenPhD · 2026-09-19
- Agents now coordinate each other — but A2A messaging sidesteps user boundaries and identity — sujingshen · 2026-09-19