Why coding agents stay dumb even as the models get smarter

Michael Lynch has a complaint that many developers will recognize. The underlying models keep getting smarter, but the coding agents wrapped around them stay frustrating. In "Why Are Coding Agents So Dumb?" he separates the two: the model is the brain, the agent is the body, and the body is where the problems live.

His examples are specific. When he asked OpenCode to add passphrase protection to his file-sharing app PicoShare, about 1,500 lines of work, it split the job into ten subtasks and ran them one after another instead of in parallel. Agents reach for an expensive frontier model to scan 50,000 lines for a pattern and never drop to something cheaper. Claude Code searched the web to learn its own features and downloaded 13 GB of files for something he never used, yet refused to ship 50 KB of its own documentation. One agent stopped two minutes into an overnight task to ask about a git branch name, then sat idle until morning. On security he is blunt: agents ignored his requests to leave certain files alone and leaked private application keys, so he built his own sandbox restricted to the repo directory rather than trust the vendors'.

Lynch doesn't have a clean answer for why this persists. His best guess is a principal-agent problem: the people building these tools optimize for demos and benchmark scores they can see, not the daily grind of the developers they may be hoping to automate away. He says he can't name the last Claude Code or Codex feature that actually improved his workflow.

Why it matters

If your coding agent feels dumber than the model behind it, the problem is usually the harness, so better results often come from switching agents or setting up sandboxing and model routing yourself instead of waiting for the next model release.