Stop Letting Coding Agents Grade Their Own Homework: A Separate Verifier Workflow
jonah_omninode · reddit · 2026-09-03
The author describes a recurring coding agent failure: completion reports sound far more reliable than the work behind them—agents run tests that don't cover the change, misread output, or claim checks passed without running them. Watching sessions doesn't fix it and doesn't scale. Their fix: the implementing agent no longer decides its own work is done. Each task starts with a definition of done containing mechanical acceptance criteria; a separate verifier inspects artifacts and gathers evidence—test results, contract and schema checks, policy gates, and readbacks from the live system—recording completion only when evidence satisfies the task contract. This proves only the specified surface, but escaped failures are recorded and turned into new tests, validators, or workflow constraints.
More from coding & agent
- Developer asks: is LangChain still worth it vs rolling your own agent harness? — curious_vii · 2026-09-03
- Inference Engineering Is Just a Recipe: vLLM/SGLang, Replicas, Cache-Aware Routing — GabGarrett · 2026-09-03
- Developer accidentally built an entire agent factory with Fable 5.1 — 0xkarasy · 2026-09-03
- Microsoft adds Fabric data agents to Foundry agents via Fabric IQ (preview) — adnan_hashmi · 2026-09-03
- Databricks pitches agent-native data infrastructure, Lakebase Postgres at VLDB 2026 — matei_zaharia · 2026-09-03
- doodlestein ships a comprehensive web app review skill after months of debugging — doodlestein · 2026-09-03