Skip to main content
An agent can call other agents to answer questions that fall outside its own knowledge. To the calling agent, this appears as an “ask another agent” tool — it poses a question, and the other agent answers. This lets you compose specialized agents into a network, each grounded in its own knowledge base, without duplicating knowledge across them.

​
How delegation works

1

The caller reaches its limit

While working on a task, the agent decides a question is better answered by a specialist.
2

It asks a callable agent

Using the “ask another agent” tool, it sends the question to one of its attached delegates.
3

The delegate answers

The delegate runs on its own knowledge base and returns a final answer.
4

The caller uses the answer

The calling agent receives only the delegate’s final answer and folds it into its own response.

​
Who an agent can call

A delegate can live in the same knowledge base as the caller, or in a different knowledge base within the same tenant — as long as that target knowledge base allows inbound sharing.
Cross-knowledge-base delegation requires the target to permit inbound sharing. See Agent sharing to control which of your knowledge bases can be reached this way.

​
Delegation is single-hop

Delegation is single-hop by design: a delegate answers the caller directly and does not itself delegate onward. This keeps behavior predictable and responses fast, because every delegated question resolves in one clear step rather than fanning out through an unbounded chain.
If you need a deeper chain of expertise, model it explicitly — attach the right specialists directly to the agent that needs them, so each hop is one you chose.

​
Reliability and access

If a delegate is unavailable or archived, the calling agent receives a graceful message and continues its work rather than stalling. Your caller stays resilient even as your set of agents evolves.
Callable agents are attached when you configure the caller, and they’re validated at deploy — so you can’t attach an agent you don’t have access to, and a deploy with an unreachable delegate won’t go through. See Deploying & versioning.
Usage from delegated agents is attributed and billed to the calling tenant, so the cost of a delegated answer lands with the tenant that asked for it.
The caller only ever sees the delegate’s final answer — not its intermediate steps. This keeps the calling agent’s context focused and its own answer clean.

​
Next steps

Attach callable agents

Add delegates when configuring the caller.

Control cross-KB sharing

Decide which knowledge bases can be reached by delegation.