Going parallel

Running a whole goal in parallel

A planner splits the goal, several coders take the pieces at once, and a reviewer reads the result.

Code
/orchestrate migrate the billing module off the legacy client

What happens

  1. A planner decomposes the goal into tasks and works out what depends on what.
  2. Coders take the independent pieces in parallel, each in its own worktree, bounded by a global cap on how many agents run at once.
  3. A diagnostics gate runs your project's checks on each result. Failing work goes back to the same agent that produced it, with the failure, rather than to you.
  4. A reviewer reads the finished set together.
  5. Each result waits for your one-key review merge.

--auto merges results that pass cleanly without stopping for you.

It survives a restart

Every group is journaled. If your machine reboots or you close the terminal mid-run:

Code
/orchestrate resume

picks the group up where it stopped. Finished work stays finished; unfinished work restarts.

When to reach for it

Good fit: a migration across many files, a rename with a long tail, adding the same handling to twelve endpoints, a dependency upgrade with scattered breakage. Work that is genuinely several tasks with a shared goal.

Poor fit: one focused change. /plan and a normal turn will be faster, and you'll be able to follow what happened.

Watching it

CtrlD shows the group. Live Grid is the useful view here — up to nine agents at once, so you can see which lane is moving and which is blocked.

Permission prompts from any agent in the group come to your prompt, tagged with which one is asking.

Cost

Parallel agents mean parallel model calls, and a large orchestration is one of the more expensive things you can do. /usage shows the spend for the session broken down by model. See Credits and what you're spending.

If it doesn't work

The plan split it wrong. Stop the group, run /plan on the same goal, fix the plan, then /a the pieces yourself. The planner is a starting point, not a verdict.

Everything is queued and nothing is running. You're at the global agent cap. Finish or stop a run to free a slot.

Two lanes conflict at merge time. Expected when tasks touch the same file. Land one, then rerun the other — with the first one landed, it re-plans against the new state.

Next