Install

Parallel coding agents: /swarm and /derby

Two commands use git worktrees to run more than one agent at once: /swarm splits one task across parallel agents, and /derby races several models on the same task.

Both need a git repository.

/swarm: split a task across parallel agents

/swarm add search, pagination and an export button
  1. A lead agent splits the task into 2-4 independent parts.
  2. Each part runs as its own agent (a "thread") in a separate git worktree, at the same time.
  3. You review the results side by side, then merge everything with Enter. Edits to the same file are combined.
  4. /undo reverts the whole swarm in one step.

Swarms work best when the parts really are independent: separate features, separate files, separate layers.

/derby: race three models, keep the best

/derby fix the flaky login test
/derby --models anthropic/claude-sonnet-5.5,openai/gpt-5.6-sol,google/gemini-3.1-pro-preview fix the flaky login test

Derby races up to three models, ideally from different vendors, on the same task. Each works in its own git worktree, so your files are untouched until you choose. Watch them work side by side, compare their diffs, test results and cost, then apply the one you'd merge with Enter.

Set your default lineup with derbyModels in ~/.darcerc (see Configuration). Browse candidates in the model directory.

Threads

Swarms are built on threads. Outside of swarms, Darce can also start sub-agents with their own fresh context on its own: several research threads can run at the same time, and a work thread can take on a self-contained change. You see each thread live, and /threads lists this session's threads (/threads <n> shows one thread's steps and report).

Next: Models · All docs · Model directory · Claude Code alternative