> ## Documentation Index
> Fetch the complete documentation index at: https://docs.cognite.com/llms.txt
> Use this file to discover all available pages before exploring further.

# About Atlas AI agent subagents

> Understand how subagents let an Atlas AI agent delegate tasks to other agents, and when to use them to build orchestrator-style workflows.

<Warning>
  Subagents (Preview) are available starting with [agent runtime version 1.3.0](/cdf/atlas_ai/references/atlas_ai_agent_runtime_versions). Do not rely on preview features in production systems.
</Warning>

Subagents let an Atlas AI agent delegate a task to another independently configured agent. The primary agent, sometimes called the orchestrator, decides when a task fits a subagent's purpose, passes relevant context to it, and uses the result to continue the conversation. The subagent runs with its own language model, instructions, skills, and tools rather than inheriting the orchestrator's configuration.

## Why use subagents

**Build an orchestrator agent.** Instead of one agent trying to handle every workflow, an orchestrator agent can route a question to an existing specialized agent, or to one you build specifically to support the orchestrator. This keeps each agent focused on a narrow, well-tested task.

**Isolate context.** A subagent runs in its own context. The orchestrator's conversation isn't cluttered with the full detail of how a sub-task was completed, and the subagent isn't distracted by instructions, skills, or tools that are irrelevant to its task.

**Match the language model to the task.** A subagent can use a different language model than its orchestrator. Route simple, well-defined sub-tasks to a faster or cheaper model to speed up the overall system, and reserve a more advanced model for subagents that handle more complex reasoning.

## Adding a subagent

Add subagents from the **Subagents** section in the Agent builder. The agent you want to delegate to must have a description.

## Approval and client-side tools

A subagent can use tools that require user approval or run client-side. When one of these tools is called during a delegated task, the approval request surfaces in the orchestrator's conversation, labeled with the subagent's name, so the user can approve or deny it without leaving the top-level conversation.

## Limits

* An agent supports up to 20 subagents.
* An agent can't delegate to itself.
* An agent can't add the same subagent twice.
* A subagent can't itself delegate to other agents. Delegation is limited to one level.

## Related information

* [Building agents](/cdf/atlas_ai/guides/atlas_ai_agent_building)
* [About Atlas AI agent tools](/cdf/atlas_ai/concepts/atlas_ai_agent_tools)
* [Agent runtime versions](/cdf/atlas_ai/references/atlas_ai_agent_runtime_versions)