Using AI Agents to Write Code

Writing all these tests myself seemed like a waste of a perfectly good AI agent, so I containerized OpenAI Codex and pointed it at the backend to write the unit test suite. This went well enough that I gave it a second job: reading Figma screenshots and implementing the React screens from them, pixel-chasing a login page and a dashboard it had never seen rendered.

The very first thing I did was hand it the keys with no supervision.

approval_policy = "never"
sandbox_mode = "danger-full-access"

Yes, that's the literal config value. Full autonomy, no review of a single command, before I'd worked out a single file permission.

That cost me the next several days. The container runs as a mapped host user via LOCAL_UID/LOCAL_GID and gosu, and getting that right took a proper tour of permission fixes. First "allow running codex shell as root," because the unprivileged user couldn't do something. Then "resolve codex permission issues," which actually fixed things by quietly removing several bind mounts (.gradle, build, just/, the Justfile) instead of solving the underlying problem. Then "allow backend codex access to more files," which added most of those same paths straight back in, plus a few more.

The container's default command was codex, meaning every container start launched a live session.

CMD ["codex"] →  CMD ["sleep", "infinity"]

The fix let the container idle until I exec'd into it, instead of spontaneously spawning a duplicate every time I started it.

The frontend copy of the agent, meanwhile, spent a while confidently reasoning about a node_modules folder that didn't exist, because nothing in its startup script ever ran npm ci. And its own briefing document telling it how its environment worked — flatly stated "network access is unavailable," which was simply untrue and had to be quietly deleted in a later commit.

prevent_repo_scan.py is a real hook that intercepts every shell command the frontend agent tries to run. It blocked find ., bare rg --files, bare fd, tree, and ls -R at the repository root, replying with a polite little message telling it to scope the search to a subtree instead. It exists because the agent kept trying to scan the entire repository before starting work on any task.

By the end of the week there were effectively three agents working on the project: one for backend tests, one for frontend screens, and a root-level one doing discovery, documentation, and final-review passes across the whole repository.