MMOOZZGG.com

Glossary

What is an AI subagent?

In short

An AI subagent is a helper agent that a main agent starts to do one part of a task. It works in its own context and returns only the result.

AI agents · subagent

MMOOZZGG.com editorsChecked: 1 min read

For a child

When Mom cooks a big holiday dinner, she asks for help: you peel the carrots, your brother sets the table, and Dad goes to the shop for bread. Everyone does their own job and nobody gets in the way. Then everyone comes back to the kitchen and tells what they did. A smart program can call helpers like that too: it gives each one a small job and gets back a finished result.

In plain words

Every model has a limited working memory. If you load a search across a hundred documents into it, it starts to get confused. So the main agent does not do the heavy work itself. It creates a helper with a short instruction and the tools it needs, for example “find all penalty clauses in these contracts”. The helper reads the documents in its own separate window and sends back a short summary. The main agent gets the summary, not piles of text, and keeps a clear head for the main task. Several helpers can run at the same time.

For business

Subagents make complex agents cheaper and more reliable: a low-cost model can do the rough work, and a strong model keeps the decisions. Each helper has its own permissions, so a document search agent can be forbidden to change files. The downside is token use and control. Anthropic reports that early versions of its research system started 50 helpers even for simple questions. Without clear rules, helpers repeat each other’s work, and the main agent does not see their middle steps. Before launch, set how many helpers are allowed, which tools they get and what summary they must return. This is how roles are shared in an AI department.

For a specialist

A subagent is an agent instance that another agent starts for a limited subtask. Its key property is its own context window, so long search results and middle steps do not fill the parent’s context: only the summary returns. In Claude Code, such agents are Markdown files with a YAML header in the .claude/agents/ folder: a name, a description, a list of tools and a model, and the body of the file becomes the system prompt. Built-in agents include Explore (read-only code search), Plan and a general-purpose agent. By default, nesting is allowed up to three levels below the main conversation. In Anthropic’s research system (13 June 2025), the lead agent usually starts 3–5 workers in parallel. The pattern belongs to the orchestrator-workers scheme and to context engineering. Its limits are synchronous waiting, no coordination between helpers and errors in the task description that the parent cannot see.

Example

A lawyer asks an agent to compare a new supply contract with the standard template. The main agent starts three helpers: one checks the terms, one the penalties and one the termination procedure. Then it merges their findings into one table.

Related terms

Questions and answers

How is a subagent different from a regular agent? It is built the same way, but another agent starts it, not a person. The helper gets a narrow task, works in a separate context and returns only the result.
Why use subagents? They keep extra text away from the main agent and let parts of a task run in parallel. You can also give them separate permissions and a cheaper model.

Sources

Where companies use it

Not sure whether your company needs this?

AI solutions map
Contact us