Documentation feedback

Help us improve this page.

About

We use submissions to improve Swarm and may reply if you include an email. Human verification is required. See our Privacy Policy.

Runtime

Agents

Swarm comes with code-owned system agents for distinct jobs. You work with Swarm; it chooses a bounded specialist when delegation will help. There are no custom agents to create, edit, or delete.

System agents

Swarm brings in a specialist when the work calls for one.

You work with Swarm directly. When a task benefits from focused research, isolated implementation, or design exploration, Swarm can delegate that part to one of three usable system agents.

Primary agent

Swarm

Scopes your request, handles cohesive work directly, and coordinates specialists when delegation will improve the result.

Coder

Implementation

Completes a defined coding scope in an isolated worktree, then returns a clean, reviewable commit for Swarm to integrate.

Finder

Research

Investigates the codebase or web, traces how things work, and returns evidence with the relevant filepaths without editing your workspace.

Designer

Design

Creates reusable design variants as managed artifacts by default, or edits an explicit workspace scope when the requested output must live in the repository.

How delegation works

Subagent work is blocking, not detached.

Finder, Coder, and Designer run as durable child sessions when Swarm delegates work to them. Swarm can launch independent assignments together, but the parent task call waits for every selected child to finish. It then reviews the reports and continues in the original conversation.

This is different from starting a separate session. A delegated subagent does not become an independent background assistant, replace the parent conversation, or continue after returning its task result. Its full transcript remains attached to its child session for review.

01 Scope

Swarm chooses a specialist only for a clear, bounded assignment.

02 Wait

The parent task blocks while the selected child sessions run.

03 Review

Swarm receives the reports, verifies the handoffs, and resumes.

Swarm is instructed to delegate when it is appropriate and likely to improve the result. You can also request delegation directly in your prompt.

Use Finder to trace how authentication works and return the relevant filepaths.

Use Coder agents to implement the independent backend and frontend changes in parallel.

Create three Designer variants for this navigation so I can compare them.

Safety boundaries

Each specialist starts with less authority than the primary.

fixed toolsEach system agent has a compiled tool contract for its purpose. Finder cannot write; Designer cannot use Bash or Git; Coder cannot use Bash, manage sessions or plans, or change user-owned todos.
no recursive delegationDelegated system agents do not receive the task tool. Child sessions are also blocked from launching subagents, so a subagent cannot create another layer of agents by default.
permission promptsIf an enabled action reaches a permission boundary, Swarm surfaces the request to you and waits for your decision. Denying it does not silently grant a fallback path.
bounded ownershipCoder works in an isolated worktree. Designer returns a managed artifact by default or receives an explicit workspace output path for repository edits. Finder remains read-only. These boundaries reduce accidental overlap and make the result easier to inspect.
durable evidenceChild sessions preserve their transcripts and final reports. Coder handoffs also retain branch, worktree, base commit, head commit, and cleanliness evidence for review.

A permission prompt is still a request to perform a consequential action, not a guarantee that the action is safe. Review it against your workspace and security requirements. See Permissions for the available approval profiles and highest-risk guidance.

No setup required

Agents are ready when Swarm starts.

You do not need to create an agent, write an agent prompt, assign tools, restore defaults, or maintain agent records. Custom-agent CRUD is not part of the current Swarm product. The runtime supplies and validates its complete system-agent registry from the installed Swarm version.

Provider, model, and thinking settings can change which model a system agent uses, but they do not change that agent's identity, purpose, prompt, mode, or tool contract. Router also specializes Iteration Swarm assignments and image briefs inside the parent-authored requirements. You can simply describe the work you want completed; Swarm handles the agent boundary.