Why autonomous coding agents burn hundreds of thousands of tokens on terminal noise, how we shrunk prompt overhead using evolutionary harness compaction, and how we built a confidence guard for the edge model that wrote this post.