Coding agents say "fixed" when they mean "I changed the code": three prompts that force verification
Ok_Negotiation_2587 · reddit · 2026-10-03
Coding agents routinely report "I edited the file" as "the bug is fixed" — they read their own edit, find it convincing, and skip actually running anything. The author shares three practical fixes:
- In the task: demand the agent run the originally failing command and paste exact output before claiming a fix; unrun claims must be marked "not verified".
- In the instructions file: ban the words fixed/works/resolved/done unless the same message shows the command run and its output. Claims get smaller and more specific, and the agent admits when it can't run something (no local DB, missing API key) instead of silently skipping it.
- End-of-session review: ask it to list every "it works / it's fixed" claim with proof or mark it unverified — usually just 2-3 items, exactly where you'd have trusted the summary.
Bonus: "run the tests" often shrinks to only the touched test file; name the full suite command when it matters.
More from coding & agent
- Meta reportedly building Agent Engine codenamed Forge for its API platform — testingcatalog · 2026-10-03
- Where should the agent end and the harness begin? A Rust runtime builder's open question — sreevarshan-xenoz · 2026-10-03
- Taskmarket launches agent labor market: fund one task, a swarm of AI agents races to deliver — 0xSammy · 2026-10-03
- Benchmarking 40 long-horizon coding tasks cost $50k in one day, pricing universities out of AI evals — himanshustwts · 2026-10-03
- The Case for Persistent Agents That Stay Connected to Work Instead of On-Demand AI — alifcoder · 2026-10-03
- Agents outpaced our org memory: a game team's workflow redesign around AI agents — EmergentInteractive · 2026-10-03