Tool errors are prompts too: adding elimination trails so agent retries can actually correct
coldoven · reddit · 2026-08-31
The author describes a core failure mode in agent loops: when a tool returns some version of "not found", the model can only retry the same call, guess another name, or answer without data — none of which is a correction, because the error never says what went wrong.
Their fix in their data layer: make resolver errors carry an elimination trail — every candidate considered and the first gate it failed, e.g. (domain): declares domain 'defaultdomain', but the run requested 'marketing'.
A remaining edge case: a typo in a feature name is reported as a missing option, which is accurate but misleading — the model adds the option instead of fixing the spelling. They haven't measured retry success before/after, and ask how others design tool errors: raw vs rewritten, and when extra detail starts to hurt.
More from coding & agent
- Intent Hailed as the GOAT of Laptop Dev Environments — Wattenberger · 2026-09-01
- Dev automated orchestration: 3 days work saves weeks of manual effort — natesiggard · 2026-09-01
- Devs debate: everyone rebuilding world-model tracking per repo — DRY failure or necessity? — deepfates · 2026-09-01
- Debunking Agent Anthropomorphism: Replaying OpenAI Incident Shows No 'Civilizations' — avlok · 2026-09-01
- Agent costs often come from pointless loops, not the model — FounderWithCode · 2026-09-01
- Snowflake research: AI Functions boost Coding Agent performance — CShorten30 · 2026-09-01