Skip to main content

Add Sub-Agents to an Agent

When to use sub-agents​

Use sub-agents when a task benefits from specialization or delegation—e.g., an HR Agent delegating ticket creation to a “Helpdesk Agent,” a Research Agent delegating web search to a “Browser Agent,” or a Planner delegating execution to domain experts. This mirrors multi-agent patterns where each agent has a tight role, tools, and knowledge.

Benefits:​

  • Encapsulates domain expertise and tools behind clear boundaries
  • Lets a “supervisor” agent plan and delegate work to specialists
  • Keeps your system modular and reusable across projects

How it works (runtime model)​

  • The parent agent runs the normal think → act → observe loop.
  • When needed, it delegates to a sub-agent, passing a single, self-contained message.
  • The sub-agent executes (using its own LLM, tools, KBs), returns results only to the parent, and the parent interprets and communicates back to the user.
  • Sub-agents do not talk to the end user directly.

Add sub-agents in the UI​

  • Open your agent.
  • In the Sub-Agents card, click Add.
  • In the Add sub-agents drawer:
    • Use the Search bar to find agents.
    • Choose My Agents to select other published agents in this project.
    • Choose A2A to select any A2A agents you’ve connected.
    • Click the + icon to attach a sub-agent.
    • Need a new one? Click Create New, publish it, then return and attach.
  • To remove a sub-agent, click the red minus icon next to it in the Sub-Agents list.
  • Preview to test delegation flows, then Publish your updated agent.

Delegation is stateless (important!)​

Sub-agents only receive the single message you send. They don’t see earlier chat turns.

Therefore, when delegating:

  • Parent agent must include every required parameter and all necessary context—no assumptions.
  • If the user references something created earlier (Agent / Tool / KB), parent agent must include its IDs and any exact values the sub-agent needs.
  • After the call returns, the parent should summarize/extract the key outcome and reply to the user.

Prompting guidance (after you add sub-agents)​

Add a short section in the parent agent’s Base instruction so it delegates reliably. For example:

Sub-Agent Delegation Rules​

  • Use Helpdesk Agent for ticket operations; include ticket_id, user_email, priority, and the exact request text.
  • Use Browser Agent for web research; include the full query, domain constraints, and max results.
  • Always package inputs as a fresh, self-contained message. Include relevant IDs if referenced in previous inputs or results for this sub-agent.
  • Sub-agents return results to you. You must interpret outputs and reply to the user.

User-prompt patterns that work well:​

Explicit delegate + payload