Part 10 of 12 · 1 min read
Quick notes
- A coding agent is a model inside a harness that gives it tools, enforces limits and decides what may merge.
- The loop: find the code, make a small exact edit, run tests, feed failures back, within an attempt budget, then a PR.
- Repository tools should be narrow: edits match exactly one place, diffs are size-capped, tests and CI stay protected.
- Passing visible tests is not correctness. Hard-coded answers and weakened tests pass them; held-out tests catch it.
- Review before merge is part of the product; the gate is a heuristic, so human review stays the backstop.