Sub-Agent
A sub-agent is an AI helper launched by an overarching AI program to handle a clearly delimited subtask, after which it reports back only its result. This breaks large tasks down into smaller portions that can be worked on separately and, in some cases, in parallel.
Some AI programs don’t just answer questions but carry out multi-step tasks: they search through files, retrieve web pages, write code, and check the result. Such independently working programs are called agents. A sub-agent is an agent that is commissioned by another agent. The overarching agent breaks the task down into subtasks and hands each one to its own helper. This helper works on its own, focuses only on its small task, and in the end delivers a brief answer back. The main agent collects these answers and assembles the overall result from them.
Why tasks are split up
An AI model can only take in a limited amount of text per request. This working memory is called the context window. With long tasks, the window fills up with search results, file excerpts, and interim notes. Important details then get buried, and the model loses the thread. Sub-agents prevent this because each helper has its own, fresh window.
The second advantage is speed. Anyone who wants to check ten sources doesn’t have to do so one after another. Ten sub-agents can start simultaneously. This costs more computing power but saves waiting time. For research that would otherwise take hours on its own, this is the decisive point.
On top of that, there is a certain degree of error containment. If a sub-agent heads in the wrong direction, this initially only affects its own subtask. The main agent can discard the result and reassign the task. The analogy is an editorial team that sends back a weak article instead of rewriting the entire issue.
Task, tools, and feedback
Technically, a sub-agent is usually the same or a smaller language model call, just with its own instruction. The main agent formulates this instruction itself: what needs to be done, what to pay attention to, in what form the answer should be given. The sub-agent is also assigned tools, such as web search, file access, or executing program code. It is often deliberately given fewer tools than the main agent, so that it cannot cause any damage.
What does NOT flow back is important. The sub-agent might read 50 pages of text but reports only a ten-line summary. All of its intermediate clutter stays in its own context window and disappears afterward. This very compression is the actual trick.
A common misconception is confusing sub-agents with Mixture of Experts. There, a model internally selects individual building blocks for the computation, invisibly and within milliseconds. Sub-agents, on the other hand, are complete, separate program runs with their own tasks. The division happens one level higher, in the software’s workflow, not within the model itself.
Sub-agents in today’s AI tools
They are most visible in research features offered under names like Deep Research. You ask a question, the system works for several minutes, and delivers a report with sources. In the background, a main agent distributes individual search tasks to several helpers. Some providers even display this in the process log.
The second major area of application is programming assistants. One helper searches for the affected files, another writes tests, a third checks whether everything still runs. Developers can partly define such sub-agents themselves and assign them fixed roles.
In business news, sub-agents usually come up in connection with costs. Each helper consumes computing time, and a task with twenty parallel sub-agents can cost a multiple of a normal request. That’s why providers limit how many helpers an agent is allowed to launch and how deep the nesting can go. Without such limits, an agent could keep spawning new agents until the budget is used up.