Run twenty agents on one Mac without a single merge conflict.

One agent at a time is your ceiling, not the model's.

• →You are waiting on a single agent to finish before the next task can start. That queue is the bottleneck, and it is not the model.
• →Everything written about parallel agents assumes cloud infrastructure and a platform bill. You want it on the laptop in front of you.
• →Without isolation two agents edit the same file and you spend the afternoon untangling it. That happens on run one, every time.
• →Twenty agents producing twenty diffs is worse than one agent unless the review cadence is designed first.

One orchestrator. Twenty workers. No merge hell.

• 01
  The topology

  Orchestrator, workers and shared memory, drawn out with the message flow between them and the point where it stops scaling.
• 02
  Worktrees as isolation

  The git primitive that makes parallel agents safe, and why every other approach to isolation costs more and delivers less.
• 03
  A one-line launcher

  The shell script that spawns a worktree, starts an agent in it, and tears the whole thing down when the branch merges.
• 04
  A review cadence

  How to take twenty incoming diffs a day without drowning, including which classes of change you stop reading entirely.
• 05
  Where the Mac gives out

  Real ceilings on M-series hardware for 5, 10 and 20 agents, and the contention signals to watch before it hangs.

You, if you are the bottleneck your agents wait on.

If you already run Claude Code daily and keep thinking there is nothing stopping five of these at once, this is the setup. An afternoon to wire, and the throughput change is immediate.

Why now Worktree-per-agent only became practical once agents got reliable enough to leave running unattended for twenty minutes. Below that reliability the orchestration cost was never worth it.

The change that lands hardest is not the throughput. It is that you stop babysitting. You queue work, walk away, and come back to reviewable diffs instead of a cursor waiting on you.