running this pattern too. one addition: having the agent write to memory files mid-session not just end of day catches more context before it gets lost
@ryancarson how are you using compound engineering with amp given that it relies heavily on custom agents and amp doesn't have custom agents as far as I know?
x.com/acolombiadev/status/20…
This is exactly the kind of setup that made me write about the maintenance problem. You’ve got an agent shipping features every night. AGENTS.md files growing. PRs stacking up.
But here’s the question I keep coming back to: who’s watching the repo health while all this code lands?
Docs drifting from code. Test coverage gaps. Dependencies going stale. CI pipelines getting slower. The stuff that doesn’t break immediately but compounds in the wrong direction.
@GitHub Next built something called Peli’s Agent Factory for this exact gap. Over 100 maintenance workflows: triage agents, CI diagnostics, doc drift detection, security monitors. Agents that maintain what your building agents ship.
I find that my biggest blocker is not choosing what to work on (which this post does a great job in tackling), but how to merge the work into main. Even with CI passes, the agent don’t always adhere to patterns, write tests, or add proof that it works.
Any advice on that front?
Hey @ryancarson, thank you for sharing your workflow in Claude Code. Followed your article and set it up successfully on Windows. I successfully ran with clauee the planning parr of your script and it returned a plan in the terminal. But when I try to run the implementation patt in PowerShell I get this error message: EBADF: bad file descriptor error on history.jsonl.lock. How can I fix it? Btw, I'm using my Max subscription and not API calls.