Update 2: The solution below includes steps to install the Pi agent harness. If you use Claude Code, this template may replace those steps.
Update: It didn't take long for this to pay off. An agent had a problem signing commits and took it upon itself to make a bunch of changes to the git config, hooks, etc. After that, any git command would crash the container. I just killed the workspace, recreated, re-ran my setup script, and was back in business - less than 10 minutes; no troubleshooting.
The idea of letting an LLM loose on my machine with --dangerously-skip-permissions has never sat well with me, but it is essentially a requirement for multi-agent workflows. So, I tried out both DevPod & Coder as a means of sandboxing agents. The latter clearly won out from a usability standpoint, but it did take some effort to get it working the way I wanted. Here's what I ended up with:
Comments
Post a Comment