Dev deletes 60,984 lines of AI-written unit tests that catch nothing
alexcovo_eth · x · 2026-09-25
Ray Fernando shared that he removed 60,984 lines of test code from his repo, arguing that latest models write unit tests that merely restate the code they were written after — they always pass, catch almost nothing, and break on every refactor, so agents waste time fixing tests instead of building features.
His cleanup playbook (adapted from Ansh): use a prompt that fans work out across parallel subagents to delete every unit test that wouldn't catch a real bug the E2E tests miss, then add rules to AGENTS.md to stop future agents from doing it again:
- Never write unit tests after you write code
- Prefer E2E tests as the sole testing mechanism for verifying complex features
More from coding & agent
- DSPy 3.4.0 adds native Jev support and ReAnchor, a new confidence-calibrating optimizer — dbreunig · 2026-09-26
- Microsoft ships a product on top of open-source OpenClaw, months-long partnership revealed — pswider · 2026-09-26
- Hugging Face Releases SmolDataEnvs: 5,000 Open RL Tasks for Small Models — marktenenholtz · 2026-09-26
- Stripe adds machine payments so AI agents can pay per API call with user identity — jeff_weinstein · 2026-09-26
- pmndrs ships Math v0.1.0: allocation-free web math engine with an AI agent skill — anselm · 2026-09-26
- Link brings purchase context to AI agents, letting 'order more supplies' just work — Aiden_Tech_Ai · 2026-09-26