Use child agents for clearly scoped Copilot Studio tasks

Visual concept: modular AI architecture and orchestration

A child agent is a lightweight agent inside the context of a main Copilot Studio agent. It can group instructions, knowledge, and tools for a single intent or task. Use one when that boundary makes the parent agent easier to understand, test, or maintain. Don’t create one simply to increase the number of agents.

What the handoff looks like

By default, the main agent sends the child agent a natural-language task. The child agent returns a natural-language summary when it finishes. You can define explicit inputs and outputs when a task needs particular values, such as a record identifier or a status to return.

The child agent receives the parent agent’s context. Give it a distinctive name and description so the orchestrator can identify when it should handle a request. Add only the tools and knowledge sources it needs. If a parent tool and child agent have similar descriptions, make their responsibilities clearer or restrict how the overlapping tool can be selected.

Example: separate a support task

For a support agent, an Order status child agent could handle order lookups, while the parent remains responsible for greeting the user and deciding where to route a request. The child agent might receive the order number in the user’s request and return a concise status summary. Test a valid order-status question, an unrelated billing question, and an ambiguous request to check that the handoff happens only when intended.

When a child agent is not the right fit

For a simple operation that needs only one tool, a direct tool or topic may be easier to maintain than another agent. Each handoff adds orchestration and execution overhead, so split responsibilities only when there is a clear boundary or a meaningful set of related instructions, tools, or knowledge.

A connected agent is a separate agent with its own orchestration, tools, and knowledge. Consider one when the work needs a separate domain or governance boundary. Its context handoff must be configured and reviewed; unlike a child agent, a connected agent doesn’t automatically receive the parent’s full conversation context. Check permissions too: a parent agent shouldn’t delegate a sensitive operation to a connected agent in a way that bypasses the intended approval or access controls.

Test the routing, not just the child agent

Test questions that should be routed to the child agent and similar questions that should remain with the parent. Check the returned answer against the tool or knowledge source, and review whether the child agent receives enough context to finish its task. If routing is inconsistent, refine the agent descriptions and instructions before adding more subagents.

Sources

• Add a child agent, Microsoft Learn
• Add other agents overview, Microsoft Learn
• Multi-agent orchestration patterns and best practices, Microsoft Learn