Defining "Done" for Coding Agents: A Task Completion Protocol in AGENTS.md
RevolutionaryBee7106 · reddit · 2026-07-27
A developer found that AI coding agents often report "done" while leaving failing tests or lint errors, causing subsequent tasks to start from a broken state.
To fix this, he implemented a Task Completion Protocol in AGENTS.md. Agents are now required to:
- Run relevant checks (tests, linting) and fix failures before reporting, or explain why they cannot
- Check if AGENTS.md or other repo instructions need updating
- Show actual results (like test duration or coverage) rather than just saying "done"
He notes that instruction priority is crucial, as injected instructions from the agent harness can sometimes override the AGENTS.md protocol.
More from coding & agent
- An AI agent audited medical-imaging papers and found widespread leakage — wandedob · 2026-07-27
- Session promises best practices for building cross-platform mobile apps with coding agents — TheOyinbooke · 2026-07-27
- Opinion: Stop Overengineering Your AI Agent Harnesses — rseroter · 2026-07-27
- A 1996 book predicted the AI coding workflow nearly 25 years early — chaitjo · 2026-07-27
- AI agents can test features instantly and turn development into a fast feedback loop — tristanbob · 2026-07-27
- Context engineering turns prompts, memory and tools into a more capable LLM stack — goyalshaliniuk · 2026-07-27