ncode Desktop / The workspace

Swarms

A swarm is a team of agents for one task. A lead agent breaks the task into parts, hands each part to a sub-agent, the sub-agents work at the same time, and the lead checks their work and reports back to you.

Starting a swarm#

Type /swarm followed by the task:

/swarm add unit tests for every module in lib/billing and make them pass

The assistant can also start a swarm on its own when a request is big enough; it uses the same machinery and you see the same cards.

Who does what#

• The lead plans and delegates. It has no file-editing tools of its own: it reads, starts sub-agents, messages them, waits for their results and verifies.
• Sub-agents do the work: they read, edit and run commands under the project's approval mode, like the assistant.
• A sub-agent can be started in the background, so the lead keeps working and gets the result when it arrives.
• Agents can message each other while they work.

Every agent appears in the agents pane with its own progress, tokens and cost; the swarm's card sums them up.

Models and effort#

Sub-agents use the conversation's sub-agent model and its sub-agent effort (/swarm_effort <level>). Pick a faster, cheaper model for sub-agents and keep a stronger one for the chat, or the other way round. See Providers, models and web search.

Limits#

Swarms are bounded so they cannot run away:

Change them under Settings → Limits. See Limits and isolation.

Isolated sub-agents#

In a git project, sub-agents can each work in their own copy of the project, so parallel edits do not collide. Their changes stay on a branch until the lead merges them. Turn this on or off under Settings → Limits → Isolate sub-agents in git worktrees.

Stopping and resuming#

Stop on the swarm's card, or /stop, stops the lead and every sub-agent. Resume starts a new lead that is told what the old workers already finished. See Watching agents work.