← Back to context

Comment by dagss

1 day ago

What are people's workflows these days?

As I use claude more and more I've started using git worktrees, one branch per worktree per PR, with possibly multiple agents working in each worktree at the same time on different aspects. And I manually instruct those agents. Like Emdash/Cursor/Zed. Sometimes I review code locally, sometimes agents push and I review in GitHub, no clear system yet. (jj seems promising, but Zed doesn't seem to support jj as well as git, so have delayed looking at it.)

But Paca is hinting in another direction where the agents are more in control of the branches/worktrees to use and are created by the agent? What tooling is used to support such flows? Would people use GitHub with Paca or is GitHub redundant as well.

This is pretty much my flow as well. Haven't gone beyond managing three work trees in parallel. It's nice being able to test locally against multiple work trees -- one is at 3000, then 3001, etc.

Have custom slash commands in forgejo that spawn an agent using deepwhale which creates a branch, then commits and opens a PR when done.

Gh issues works surprisingly well as an agent board. Labels for state, one issue per feature. The part i haven't figured out yet is how to know when the output is actually done vs just "looks done" to the agent.

  • Been thinking about this. "Done vs looks done" is partially who is accountable for calling it done, and the trap is the agent that did the work also often declaring it done.

    Cheapest fix: a separate done-caller (another agent or you) against criteria written before the work. Reviewer is never the author. (Basically RACI, responsible != accountable)

  • I find well described but concise acceptance criteria does a good job of anchoring the llm to the correct output. Also have them take screenshots of any UI work and respond to the ticket with them as proof.

How much are worktrees benefitting you? If I can describe the work so clearly that it can be done in parallel, I find Claude can typically one shot so parallel work isn't needed.

  • Not sure what you mean. Whether Claude one shots or not, it spends some minutes and in those minutes I can start another task in another worktree...