TRIM attacks coding-agent code bloat

The TRIM paper studies agent-generated code bloat and frames the culprit as the agent search trajectory, proposing trajectory minimization as a way to reduce oversized implementations. Why it matters: Code size is a review cost, not just an aesthetic problem. If agent loops reward long detours, teams need evals that measure minimal acceptable diffs instead of only pass/fail task completion.

Try this: Add one review metric to coding-agent trials: how many files, lines, and intermediate steps changed before the tests passed, compared with a human baseline.