AI Compass / Quick reads
The check rules do not belong to the agent
If you let an AI improve itself, you have to hold the yardstick fixed. Otherwise the yardstick gets improved.
In depth: Self-improving loop with locked checks
If you let an agent improve itself, it likes to find the shortest path: it changes the measurement instead of the result. AI LABS transfers Karpathy’s AutoResearch to Claude Code. The checks are written, approved by you, moved into a locked folder and committed. After each round, the agent may only rewrite the “How to work” section of its instructions program.md. According to the AutoResearch README, the evaluation stays unchanged there too (externally checked). Two lessons from the video: only one feature per loop, and green checks only say something about what they check. All 11 were green, yet the order form had not been tested at all.
Tool and alternative
All you need is Claude Code and git. The deny rule in the settings is not a watertight wall: according to the docs, a script that opens files itself bypasses it (docs checked). The finished skills from the video are in the paid community.
Sources
- AI LABS, “He Finally 10x Claude Code With This Method”, 2 October 2026 (3:30, 5:24–5:56, 8:09, 8:42–11:56)
- karpathy/autoresearch, retrieved 17 October 2026
- Claude Code docs Permissions, retrieved 17 October 2026
Advertising in the source: Upstash Box (approx. 6:29–7:35), plus a paid community.
Want to try this with your team? That’s what the AI Lab is for. See the AI Lab