Skip to main content

Overview

Sub-agents are specialized agents you can delegate to. Each runs in its own isolated context with a restricted toolset, so the main conversation stays focused while a sub-agent handles a self-contained task — exploring the codebase, reviewing a change, or any role you define.
Delegate by @mentioning a sub-agent in the composer, or by using one as a node inside a Workflow. A sub-agent’s full transcript opens in its own detail window.

What You Can Configure

Create and edit sub-agents in Settings → Extensions → Agents.
FieldDescription
NameCanonical kebab-case identifier (e.g. codebase-analyzer)
Display nameHuman-readable label shown in the UI
RoleOne-line description shown to the main model
System promptThe instructions that define the sub-agent’s behavior
Allowed toolsThe tool allowlist for this sub-agent
PermissionsPer-capability allow / ask / deny overrides
ModelOptional model override (otherwise inherits the parent’s)
Max iterationsCap on the sub-agent’s steps per delegation
Icon & colorOptional SF Symbol and accent color
Sub-agents cannot dispatch other sub-agents — delegation is one level deep, which keeps runs predictable.

How Delegation Works

1

Create the sub-agent

Define its role, system prompt, allowed tools, and permissions in Settings → Extensions → Agents.
2

@mention it

In the composer, @mention the sub-agent to hand it a scoped task.
3

It runs in isolation

The sub-agent works with its own tools and context, then returns a result to the main conversation.
4

Inspect the transcript

Open the sub-agent detail window to see exactly what it did, step by step.
Keep each sub-agent narrow — a tight role and a minimal toolset produce more reliable results than one do-everything agent.

Workflows

Compose sub-agents into multi-step graphs with evaluators, breakpoints, and checks.