Why Are Coding Agents So Dumb?

Why we picked it · the editor's summary

Asked to add passphrase protection, about 1.5k lines, OpenCode split the work into 10 subtasks and ran them one after another, and Claude Code sits for minutes waiting on end-to-end tests before it starts the commit message. Lynch, who uses both daily, lists more: agents never hand tasks that are 95% gruntwork to a cheaper model, search the web to learn their own features, write plans as piles of low-level details, and stop an overnight task two minutes in to ask for a branch name. Asked politely not to read certain files, his agent sent private application keys to OpenAI and Anthropic, so he built a sandbox that keeps agents inside the repo directory. His best guess is a principal-agent problem: demos and benchmarks measure models, not agents. If you run coding agents, enforce file access at the OS level, not in the prompt.