Each agent runs in a separate worktree or cloned copy of the repo, independently. When the task is done, a PR is opened. The issue you mentioned gets caught during merge conflicts.

No different than a "real" dev team?

Ideally, if you are able to use code to validate (either with a test or eval) that works best.

I had a huge number of database rows that each needed LLM reasoning and tool use.

Existing harnesses like Claude Code don't handle this well - give it a batch and it'll try to write you a script instead of actually reasoning over each item.

Even when it works, context bloats after each turn - gets expensive, slow, and starts hallucinating. And there's no built-in retry handling or parallelization.

So I built this tool. Define the task as a prompt, pipe in your task list, set concurrency, model, tools, etc.

Open to feedback!

Comments URL: https://news.ycombinator.com/item?id=47507943

Points: 2

# Comments: 2