Sub-Agent Collaboration

Mechanism

The main agent creates sub-agents for sub-tasks:

Independent sessions and contexts per sub-agent
Main agent aggregates results
Right panel “SubAgents” tab shows real-time status

Model Tiering

Settings > Models > Sub-agent tiers:

Tier	Best For
Simple	Retrieval, formatting, validation
Standard	Daily coding, writing, analysis
Hard	Complex reasoning, architecture, multi-file refactoring

Unconfigured tiers follow the parent session model.

Sub-Agent Discipline

Sub-agent tool calls follow the sub-agent approval policy (Settings → Security → Sub-agent approval)
Follow parent session (default): approval cards show in the main chat’s approval area, tagged with the originating sub-agent, at the same approval level as the parent
When off, routine sub-agent operations auto-approve (no dialogs), while dangerous operations still require confirmation

Two guardrails hold in every mode: dangerous operations always require confirmation, and sub-agent deletion is always forbidden.

Sub-Agent Roles and Tool Whitelists

Read-only roles can be assigned when dispatching sub-agents:

Role	Purpose
`explore`	Read-only explorer, small-model tier by default; great for mapping code structure
`review`	Read-only code review

You can also restrict which tools a sub-agent may use—give it only what it needs. Its permissions never exceed the main session’s.

Worktree Isolation and Snapshots

Coding sub-agents run in their own worktrees, so their changes never touch your main directory. Output is snapshot-protected and recoverable even after cleanup. See Git Worktrees.

Contexts preserved in session records after completion

Example

Do three things simultaneously: 1) Research competitor pricing 2) Organize monthly expenses 3) Write a product update announcement

Three sub-agents execute in parallel; the main agent aggregates.

External Sub-Agents

Supports external agents (Codex / Claude Code) via ACP protocol.