Four worktrees, four agents

Many agents, one playbookDeveloper running agents in parallel git worktrees

We ran four agents in four git worktrees on separate tasks. Two of them took the same follow-up task.

How do I stop parallel AI agents in different git worktrees from doing the same task?

Run arcroom init --everywhere so ArcRoom covers every clone and worktree of the repo. Agents claim tasks in the shared room, so a task already claimed is skipped, and each finished task is reused by the others. Claims only help if each agent calls claim_task.

Also asked as: Two of my parallel agents picked up the same follow-up task. How do I prevent that? · How do I share memory across git worktrees for AI agents? · Can AI agents in separate worktrees see what the others have claimed?

The work

arcroom init --everywhere covers every clone and worktree of the repo.

ArcRoom keeps
  • Reused answer
  • Claim
The next agent gets

[worktree 3] claim_task → already claimed in worktree 1: "update API docs for search"

Step by step

  1. arcroom init --everywhere covers every clone and worktree of the repo.
  2. Agents claim tasks in the shared room, so a claimed task is skipped.
  3. Each finished task is reused by the others.

What the agent sees

[worktree 3] claim_task → already claimed in worktree 1: "update API docs for search"

Compared with today's tools

A task list with owners (TODO.md or issues)

Give each worktree's agent one issue, and keep follow-ups in a list a person hands out.

Setup
15 min
Ongoing
a person dispatches every follow-up

ArcRoom

Every agent reads the team playbook before it works and writes to it when it finishes. Nothing to update by hand.

Setup
2 min (npx arcroom init --everywhere)
Ongoing
None

Human dispatch avoids duplicates just as well; ArcRoom lets agents claim follow-ups themselves without someone watching.

Where it stops

Claims help only if each agent calls claim_task.

Try it in your project

Works with the agents your team already uses, and any agent over MCP or the HTTP API.

More in many agents, one playbook