Multi-Agent Teams
DocsUse WindOpMulti-Agent Teams6 min read
Documentation

Multi-Agent Teams

Split big jobs across specialist agents. Each one gets a focused goal, runs in isolation, and reports back so your main chat stays clean.

Delegating work

The agent can hand tasks to sub-agents with the delegate_task tool. You do not need special syntax — describe the split you want.

prompt
Have three agents review src/auth.ts in parallel:
1. a bug hunter
2. a security reviewer
3. a performance reviewer
Then summarize the findings by severity.
  • Up to 20 independent tasks per request, with 3 workers running at once by default and a hard maximum of 5.
  • Each task can override the role, toolset, and model.
  • Every sub-agent keeps a transcript you can read back with get_subagent_transcript.

Execution styles

StyleHow it worksBest for
ParallelIndependent workers run at the same time; results are merged.Multi-angle review, research across sources
SequentialEach step builds on the previous output.Pipelines: research → plan → implement → review
DiscussionAgents respond to each other over rounds; a supervisor synthesizes.Trade-offs and architecture decisions

Team and group chats

In a group chat, agents take turns; routing can be round-robin, random, or auto (a supervisor picks who should answer).

Pick which model each role uses under SettingsModels → Team & Group Chat Models. Roles: Code, Research, Review & Supervisor, Vision, and General.

Kanban board

For larger projects, the agent can plan work on a board and dispatch tasks to sub-agents.

ColumnMeaning
Backlog / To doNot started
In progressAn agent is working
ReviewWaiting for checking
DoneFinished, with result
BlockedWaiting on something else

Tasks have priorities (critical, high, medium, low), assignees, tags, dependencies, and optional iterations. See the kanban_task and kanban_dispatch tools in the Tools Reference.

Tips

  • Give roles specific expertise — “security analyst focused on OWASP Top 10” beats “security person”.
  • Use parallel for independent work; sequential when steps depend on each other.
  • Each agent is its own model call: three parallel agents cost roughly three times one. Start with two or three.
  • Track spending in SettingsUsage.

What's next

Scheduler & Automation

Run agents on a schedule.

Tools Reference

delegate_task and Kanban tools.

Something missing or out of date? Tell us on the community page.