Delegated agent task

Hands a focused task to a separate AI agent with its own instructions and tools, and returns the result to the run.

When a step needs isolated context or deeper reasoning than the main run, like validating documents against specialized compliance rules, or extracting structured data with a strict schema. Used sparingly; most steps don't need it.

What the architect sets up

• The task instructions and how capable a model it gets
• Which of the workflow's tools it may use
• The expected shape of what goes in and what comes back

How it behaves in a run

• Runs when the agent decides it's needed, or automatically when a stage begins or completes.

Related

• Stages & components: how components fit into a workflow
• Showcase: see components like this at work in real workflows

For developers: the API and MCP server identify this component as subagent_delegator.