Delegating to a Subagent

Once you have one or more subagents, the main chat can hand off work to them. The delegating agent sends a task, the subagent runs it with its own tools and prompt, and the result comes back as a reply you can keep working with.

This guide covers two ways to delegate:

  • Implicitly, by letting the main chat decide.
  • Explicitly, by telling the chat which subagent to use.

Before you begin

  • At least one subagent must exist in your workspace.
  • Its Description should clearly state what it does. The main chat reads the description to decide whether to delegate. Agents with vague descriptions get ignored.
  • You must be in a workspace where you have access to the agent (owner, shared with you, or workspace admin).

Implicit delegation

Open the main /ai chat and ask for something the subagent can handle. If the request clearly maps to one of your agents, the chat will call the agent on its own.

Example:

"Give me the support digest for yesterday."

If you have a subagent called Support Digest with a description like "Produces a daily summary of support tickets grouped by severity," the main chat will notice, delegate to that agent, and return the summary.

Implicit delegation works best when:

  • You have a small number of subagents with clearly non-overlapping descriptions.
  • The request uses words that echo what the subagent is for.
  • The agent lives in the same workspace as the chat.

Explicit delegation

Name the agent directly:

"Ask the WMS Drift Monitor to check for any new discrepancies since this morning and give me the list."

"Use the Weekly AR Reminder subagent to draft the email for the three oldest invoices in the report you just pulled."

Naming is the safer choice when:

  • Two subagents could plausibly handle the same request.
  • The work is sensitive and you want to be sure the right guardrails apply.
  • You are building a multi-step workflow that hands off to specific agents at specific steps.

Chatting directly with a subagent

You do not have to go through the main chat at all. From the Subagents page, click the chat icon on an agent's card. This opens /ai with that agent pre-selected. Everything you send goes to that agent, with its system prompt and its tool whitelist — no routing step.

Direct chat is the right mode when you are:

  • Developing and iterating on the agent (refining its prompt, adding tools).
  • Doing focused work that only that agent should touch.
  • Running a sensitive operation where you do not want the main chat to second-guess which agent to use.

Chaining subagents

A subagent can delegate to another subagent. This lets you build small pipelines: a "Report Builder" agent that delegates data gathering to a "WMS Query" agent and formatting to a "Markdown Formatter" agent.

Chaining is bounded by a delegation depth limit so agents cannot recurse forever. By default, a subagent called by the main chat cannot delegate further. If you need deeper chains, ask your workspace admin to raise the limit.

What happens during delegation

  1. The main chat (or calling subagent) picks a target agent.
  2. It sends a task string and, optionally, context from the current conversation.
  3. The target agent runs with its system prompt, its model, and its tool whitelist — not the caller's.
  4. The target agent returns a text result to the caller.
  5. The caller continues the original conversation using that result.

The delegation is a sub-run. You see it as a tool call in the main chat, and you can expand it to read what the subagent did.

Next steps