Subagents

A subagent works on part of a task in its own conversation and reports back to the agent that started it. The delegation plugin, which the default agent carries, gives an agent three kinds: a copy of itself, the read-only explore, planner and review roles, and forks of its own work.

The agent hands a child work that stands apart from what it is doing, or that would fill its context with material it needs only a report of. The agent decides when to delegate. You can also ask for it in your own words.

Asked for a reading guide to a folder of design documents, the agent split the reading across ten explore children in one call. Each Subagent link opens that child. Click to enlarge.

A copy of the agent

When the agent delegates without naming another agent, the child is a copy of the delegating agent: the same plugins and the same model, unless the agent picks another model for it. A child of a domain agent keeps that domain’s tools and skills.

The explore, planner and review roles

The delegation plugin also brings three roles. Each runs with the delegating agent’s plugins, read-only: any tool that would change something is refused, with the reason “This session is read-only”.

Role	What it does	What it returns
explore	Surveys files, data, task records, logs and sources.	Findings, each with where it came from, such as a path and line, a task, a log entry or a URL, and where it looked and found nothing. No proposals.
planner	Reads what it needs and designs a workflow for the assignment.	A workflow the delegating agent can start unchanged, and a few sentences on why these steps.
review	Checks work or a claim it did not do against the material it rests on.	A verdict: it holds, it holds with the caveats named, or it does not hold, with what was checked for each point.

A read-only role cannot start subagents or workflows, and it does not ask you questions; it reports open choices in its result. The agent can also delegate to subagents its plugins define, or to any installed agent.

What a child starts with

A child starts with only the task the agent writes for it. It does not see the parent conversation: not your messages, and not the agent’s earlier work. The delegating agent is therefore asked to write a self-contained task, with the goal, the inputs and where they are, the constraints, and the report it wants back. Inspect the assignment to read exactly what a child received.

A child is told that no person reads its conversation, to stay inside its assignment, and to end with a report the delegating agent can act on: what it found or did, the evidence, and what remains open. It does not receive the guidance for presenting results to you.

A child that is not read-only can still ask you a question. The question appears in the conversation that started the work, named after the child, and your answer goes back to the child.

Forks

A fork is a copy of an agent’s work so far, sent in another direction with a note that says what to do differently, such as testing a competing explanation. The agent decides when to fork. A fork runs on the same model as its source.

Where	What the fork is	Read more
A conversation	A new conversation with a copy of the history, listed under its source in the left panel as Forked by the agent. The note is its first message.	Forks the agent makes
A subagent	A copy of the child beside it under the same parent, named after it with ~1 added. The parent’s wait covers both.	Subagent controls
A workflow step	A copy of the step drawn beside it, b~1 for step b. The steps that waited on the step read both answers. Workflow approval and growth limits apply.	Fork a step

Forks per task under Settings → Execution & history → Subagents limits how many forks one task, or one conversation, may have. Zero means unlimited.

See them work

Two recorded conversations started from a plain request that named no tool or role; the agent chose to delegate:

• A reading guide: ten explore children read a folder of 93 design documents, and the agent wrote the guide from their reports.
• A skill review: six review children checked every plugin’s skills, and the agent merged their verdicts.

A branched study shows copies of the agent playing a study’s games in parallel. Real conversations walks through the recordings, including an agent fork.