| name | agent-swarm |
| description | Agent Swarm — parallel multi-agent orchestration. Use when a task has several independent parts and would go faster delegated across specialists (researcher, analyst, writer, reviewer, optimizer, designer, builder) than done step by step. |
Agent Swarm
Instead of one agent working through a task in sequence, Agent Swarm splits it
into independent units and runs specialists on them at the same time.
When it helps
- The task has genuinely separate parts — different subjects, sources,
sections, or files.
- You want several lenses on the same thing — draft it, check it, tighten
it, and think about how it is presented.
- Gathering can happen while something else is being produced.
When it does not
A single small change is faster done directly. Every specialist is a separate
run, so a swarm on a one-line task costs more and returns later for no gain.
Match the number of agents to the number of real units.
How it works
- Decompose — split the request into units that do not depend on each other.
- Delegate — hand each to a specialist with the
task tool, briefing it
properly: goal, context, constraints, what to return. It starts with none of
this conversation.
- Parallelise — send the independent units in one message, as several
task calls at once.
- Iterate — feed results into the next wave (produce → check → revise).
- Synthesise — merge into one answer, resolving disagreements between
specialists rather than pasting both.
Specialists
| Agent | Use for |
|---|
researcher | Finding things out: sources, options, prior art |
analyst | Data: inspecting, computing, concluding |
writer | Long-form prose: drafting and editing |
reviewer | Correctness, gaps, unstated assumptions |
optimizer | Smaller, faster, cheaper, clearer |
designer | How the result is presented and structured |
builder | Making things: code, configs, scripts, files |
plan | Sequencing work and finding dependencies |
explore | Fast orientation in an unfamiliar codebase |
general | Anything without a better fit |
Enter it by picking swarm from the agent list (Tab), or by asking for the
work to be split up.
