engineer in Finland. the production half of agents (course + book ↓) and https://teapaw.org/t.co/Ld3IjLVBlH, a daily minimum for writing in a language
claude code · mcp · custom skills
An agent reviewed 12 pull requests before any human opened them. 1 merge, 3 holds, 8 blocks, each with a reason, a line number and the test result on the patched tree.
$0.70, 2 min 11 s, 10 of 12 verdicts match the answer key.
8 modules, 55 lessons, 65 stages graded by tests. Free, runs locally, the model through your Claude Code login, no API keys. Every number above comes from a committed, dated run.
factory.lihachev.pro/?utm_so…
The weak spots. Precision over all findings is 0.58 (19 findings against 12 planted labels), so about 4 in 10 are noise someone still has to dismiss.
On blockers precision is 0.875 and recall 0.8: it caught 4 of the 5 merge-stoppers and missed one.
Every number sits in a committed record with a date and a model id (claude-opus-5-5, Agent SDK). The label file is never read during a review.
It's the capstone of a free course on the production half of agent systems, lessons graded by tests: factory.lihachev.pro/m7/?utm…