Subagents
Split big tasks across specialists. An agent, the parent, can hand part of a task to another agent, its subagent, and use the result. The parent decides on its own when to delegate.
A subagent can be:
- An agent you built, such as a researcher or a reviewer.
- An external agent on another system, over A2A.
Add Subagents
- Open the parent agent and go to Tools, Skills & Knowledge.
- Under Subagents, choose Manual and select the agents.
- Save, and send a task in Chat that needs one of them.
What to know:
- All agents lets the parent agent use any agent the person can access, including agents created later. To leave one out, click Exclude a local agent. External agents are never added this way.
- External agents are added after the parent agent is created. See External Agents.
- Subagents must share the parent agent's environment.
- Automated runs, with no signed-in person, use only the agents you selected.
- For long coding tasks, pick a subagent with a dedicated runtime.
Fork the Agent
Keep long searches and noisy tool output out of the conversation. Every agent can hand a task to a fresh copy of itself, and only the copy's answer comes back. You don't need to set this up.
The agent picks the model the copy runs on for each task:
- Normal: the agent's own model. This is the default.
- Smart: a stronger model, for hard reasoning or checking a conclusion.
- Fast: a cheaper, quicker model, for bulk reading, searching, or extraction.
To choose the Smart and Fast models, set Fast model and Smart model next to the default model in Settings → Agents. They apply to every agent, can be on any provider, and are separate from an auto router's own fast and smart tiers. A Normal copy of an agent on an auto router is routed by that router. Unset tiers use the provider's latest small model for Fast and its latest flagship for Smart.
What to know:
- The copy has the same tools and instructions, but none of the conversation. The agent writes everything the copy needs into the task.
- The copy uses the agent's key when that key serves the tier's model, and otherwise a key the person may use for it. With no such key, it uses the provider default for the tier on the agent's key, or else the agent's own model.
- A copy cannot fork again.
Get Trusted Answers From Untrusted Data
Let a subagent read untrusted data, such as inbound email or a web page, without the parent losing trust. With Guardrails on, a parent can ask for a structured answer, such as a number of days or one choice from a fixed list. An answer that matches comes back at the parent's own trust.
What to know:
- The parent names the answer's shape when it delegates. Every field must be bounded: a number in a range, a yes or no, or a fixed list of values. Free text is never accepted.
- An answer that does not match is withheld. The parent gets an error, never the raw text.
- External agents cannot return structured answers.
Answers the Parent Must Accept
A subagent can read a web page or an inbound email and finish its task, and the parent keeps its trust until it chooses otherwise. With Guardrails on, an answer that would lower the parent's trust or narrow its audience does not come back on its own. Guardrails holds it, and the parent picks one:
- Accept: the parent gets the answer and takes on its labels. After accepting an answer built on an untrusted web page, the parent is untrusted too.
- Discard: nothing of the answer reaches the parent, and its labels stay as they were.
To keep the parent's trust and still use the result, ask for a structured answer instead.
What to know:
- An answer that changes nothing for the parent comes back directly.
- A discarded answer cannot be accepted later. The parent delegates again to get a new one.
- This covers forks and agents you built. The exceptions are a subagent with a dedicated runtime and a subagent's own subagents: they are refused such a read instead.