Sub-Agent

What is a sub-agent?

A sub-agent is an AI agent assigned a bounded part of a larger task by a parent agent or orchestrator. It usually receives its own objective, relevant context, instructions, tools, and expected output, then runs a separate decision loop.

The result returns to the delegating system for review or integration.

How sub-agents are used

Suppose a primary agent must prepare a product launch report. It may delegate competitor research to one sub-agent and risk analysis to another. Each receives only the context needed for its assignment and returns a structured result with sources.

The parent or orchestrator remains responsible for coordinating dependencies, resolving conflicts, and producing the final deliverable.

Sub-agent versus tool

A tool performs a defined capability when called, such as querying a database or calculating a value. A sub-agent can make multiple decisions, use tools, react to observations, and adapt its approach while pursuing a delegated goal.

Delegation is therefore different from tool calling. However, an application may expose a sub-agent behind a tool-like interface. The implementation should still define whether the called component is allowed to act autonomously.

Sub-agent versus workflow step

A workflow step follows predefined logic. A sub-agent has its own agent loop and can choose among actions within its boundaries. Not every step needs a sub-agent; deterministic processing is often simpler and more reliable.

A sub-agent also differs from an ordinary model call. One call may generate text without maintaining state or taking actions.

Delegation requirements

A useful sub-agent assignment should include:

• a specific outcome and completion test;
• relevant context without unnecessary data;
• allowed tools and permissions;
• format and provenance requirements;
• time, cost, and retry limits;
• escalation behavior when blocked.

Poorly scoped delegation creates duplicated work, missing assumptions, and incompatible outputs.

Risks and boundaries

Sub-agents can amplify errors if the parent accepts results without validation. Handoffs may drop important constraints, and parallel agents can act on stale shared state. Separate permissions and logging help contain risk.

The parent should distinguish verified observations from a sub-agent’s interpretation and preserve source evidence.

Why it matters

Sub-agents can add specialization, isolation, or parallel execution to complex tasks. They are valuable when a delegated problem is substantial enough to justify an independent agent loop.

For small deterministic work, a tool or normal workflow step is usually clearer. See [Single-Agent vs Multi-Agent Systems](/single-agent-vs-multi-agent-systems/) for when delegation improves a system.