4 of 5 tampered changes pass a green test suite — can you trust your agent's 'done'?
Weary-Information-15 · reddit · 2026-10-21
The author ran an experiment: after an agent produced a clean test run, they introduced five changes to a repo. Four left the suite green — a weakened assertion, a new root conftest.py, and a pytest.ini that collects but runs nothing all passed; only breaking the implementation turned it red.
Key point: this isn't a clever attack — it's what happens when the thing being measured can reach the measurement. When agents can edit tests and configs, exit code 0 means little.
The author asks what people actually do before believing an agent's "task complete" — suspecting that most just merge on the green check, and few can name a check that has ever caught anything.
More from coding & agent
- A simple context hygiene trick for Claude Code: audit CLAUDE.md with claude-md-management — daniel_mac8 · 2026-10-21
- Laude Institute backs WaggleHaus, an open-source query engine for agentic queries over data lakes — lateinteraction · 2026-10-21
- komorama's weekly: the AI Peter Principle—first 90% never faster, second 90% never slower — lennysan · 2026-10-21
- Orchestrating open-source video models with Opus 5.5 cuts costs ~80x vs Seedance — mickeyxfriedman · 2026-10-21
- Multi-agent swarm architectures are being discovered, not designed — shared state emerges as the norm — JensHonack · 2026-10-21
- UC Berkeley open-sources SkyDiscover, an AI framework that discovers algorithms — CShorten30 · 2026-10-21