Michael, the post author here. I build Cmd+Ctrl, a unified mobile app for all the AI coding agents.
I typically run a single orchestrator agent across 10–15 parallel worktrees. When the orchestrator says one is ready for me, I'd like that to actually be true.
The change that has made the biggest difference for me is rather than asking independent agents to REVIEW the code, I tell them the code is broken and to prove it, and I do it in a loop. It's been very effective.
I'm curious to know whether others have utilized this approach or what other successful approaches people have hit upon to get their coding agents as far as they can without needing human review and validation.
Michael, the post author here. I build Cmd+Ctrl, a unified mobile app for all the AI coding agents.
I typically run a single orchestrator agent across 10–15 parallel worktrees. When the orchestrator says one is ready for me, I'd like that to actually be true.
The change that has made the biggest difference for me is rather than asking independent agents to REVIEW the code, I tell them the code is broken and to prove it, and I do it in a loop. It's been very effective.
I'm curious to know whether others have utilized this approach or what other successful approaches people have hit upon to get their coding agents as far as they can without needing human review and validation.
I've packaged the prompt and the loop as a skill on GitHub: https://github.com/cmdctrlai/refute