subagent

advancedspreading

A subagent is an assistant agent launched by another AI agent for a well-defined subtask, which returns only the result at the end.

In plain terms

Imagine you are renovating your apartment and hire a general contractor. They do not do everything themselves: they subcontract the electrical work, tiling, and painting, then take over the finished work from them. A subagent is such a subcontractor in the AI world: a "main" agent calls it in for a subtask.

This is necessary because an AI agent's working memory is limited. If a single agent read a hundred documents, the many details would push out of its attention what you originally asked for. The subagent performs the detailed work in its own, separate working memory and returns only a concise summary.

The other advantage is parallelism: the main agent can launch multiple subagents at once, so what would take an hour sequentially can be completed in a few minutes. Meanwhile, you typically only converse with the main agent, while subagents work in the background.

How it works

The main agent assesses the task and decides which part to delegate, for example, one that can be done independently or involves extensive reading.
It writes a briefing for the subagent: what the goal is, what it must return, and what to watch for. The subagent knows only this much about the task, as it usually does not see the previous conversation.
A subagent starts fresh, with its own working memory and authorized tools (such as search or file reading), and operates independently.
At the end, it sends a brief report back to the main agent; the material and details read in the meantime are not transferred.
The main agent compares incoming reports, verifies them if necessary, and compiles the response for you from these.

An everyday example

You ask your AI assistant to compare five home insurance policies. The main agent launches five subagents, each reading through the conditions of one insurer, and then the main agent creates a table for you from the five summaries received. You only see that the answer arrives surprisingly quickly and thoroughly.

Who is affected

developerCoding agents search the codebase with subagents, run tests, or perform code reviews, so the developer must understand when it is worthwhile to use them. → Code

entrepreneur, executiveSubagents speed up work but incur multiple computing costs, making it a business decision on when it is worth deploying them.

employeeResearch and office AI tools increasingly work this way in the background, so it is good to know that the response received is a summary of multiple agents' work. → Work & HR

A common misconception

MythThe more subagents working on a task, the better the result will be.

In realityEvery handover can lose information, and every subagent incurs a separate cost. Many tasks are solved more accurately and cheaply by a single agent; a subagent is useful only where the subtasks are truly independent of each other.

Limits and risks

A subagent only knows what the main agent described in the briefing, so incomplete or ambiguous instructions lead to poor results.
The returned summary sounds just as confident even if it is wrong, and the main agent does not always detect the error.
Running multiple subagents in parallel multiplies computing costs, which quickly shows on the bill or usage quota.
During work, it is harder to track exactly what each subagent did, so for sensitive operations, it is important to know which tools and data they have access to.

Similar, but not the same

multi-agent system – A multi-agent system is the broader concept where agents can cooperate as equals, while a subagent is always subordinate: it is launched by another agent and reports back to it.
tool calling – When calling a tool, the agent uses a ready-made function (for example, a search engine or calculator), whereas a subagent is itself a thinking agent that solves a task independently in multiple steps. In practice, the two merge: the main agent usually launches the subagent via a tool call as well.
orchestrator – The orchestrator is the main agent that assigns and coordinates the work, while the subagent executes the assigned subtask.

Related concepts

Broader

AI-agent, multi-agent system, agentes AI

Narrower

research subagent, code-review subagent

On the edge

orchestrator, context window, tool calling, coding agent, prompt

A brief history

The idea that a complex task is solved by many smaller, cooperating agents is old: Marvin Minsky's 1986 book, The Society of Mind, described human thought as the cooperation of many simple 'agents' that are not meaningful on their own. Experiments with agents built on language models have been conducted since 2023, and subagents became widely known in 2025 when they appeared as built-in features of coding and research agents.

Where things stand

As of late 2026, subagents are a standard part of coding and research agents, and developers use them daily. The average user typically encounters them without knowing it: the AI assistant distributes the work in the background.

Status: 2026 November

Also known as

subagent · sub-agent · alagent · al-agent · helper agent · sub-agent · worker agent · child agent

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.