A subagent is a separate Claude instance that a main agent (often called the coordinator) starts to handle one focused task. It begins with a blank page: it knows only what the coordinator puts in its prompt. That isolation keeps each task clean and focused, but it also means the coordinator must pass over everything the subagent needs, explicitly, every time.
What a subagent actually is
In an agentic system, one agent can do a lot of work in a single long conversation. As the work grows, that conversation fills with search results, file contents, failed attempts and side notes. A subagent is one way to avoid that. The coordinator hands a self-contained job to a new instance, the subagent does it in its own context, and a result comes back.
The glossary defines it briefly: a separate Claude instance spawned for a focused task, starting with only the context it is given. The coordinator is the agent that splits the work, dispatches subagents and merges what returns.
Why context isolation helps
Isolation is a feature, not a limitation. Three benefits stand out:
- A cleaner working context. The subagent reads many documents or runs many tool calls, then returns a short summary. The coordinator never has to carry all that raw material.
- Focus. A subagent with one job, one instruction set and a small set of tools is less likely to be distracted than a single agent juggling every tool and every task.
- Independence. Tasks that do not depend on each other can run side by side instead of one after another.
Think of a manager delegating to a new contractor. The contractor is capable, but has not attended any of your meetings. Whatever they produce depends on the brief you give them.
The catch: a subagent knows only what you tell it
This is the idea people most often get wrong. A subagent does not inherit the user's original request, the coordinator's earlier findings or the results of previous tool calls. It does not remember an earlier subagent either. If a detail is not in its prompt, it does not exist for that subagent.
So a good brief states, in the prompt itself:
- The task, in one or two clear sentences.
- The material it needs, such as the relevant text, data or file locations, rather than a reference to "the document we discussed".
- The output format you want back, so the coordinator can combine results without guesswork.
- The quality bar: what counts as done, what to flag and what to leave out.
A vague brief such as "check the contract" is the classic failure. The subagent has no idea which contract, which concerns or what to return. Paste or point to the actual contract and say what to look for.
Running subagents in parallel
When tasks are independent, the coordinator can start several subagents in the same turn rather than waiting for each to finish. For example, a coordinator preparing a market briefing might send one subagent to summarise customer feedback, another to review competitor material and a third to check a set of figures. None needs the others' output, so they can run at the same time, and the coordinator merges the three results.
If one task depends on another's output, it has to wait. The coordinator should start it only after the earlier result is back, and include that result in the new brief.
What the coordinator is responsible for
Delegating does not mean giving up responsibility. The coordinator:
- decides how to split the work;
- writes each brief completely;
- merges the results and notices gaps or contradictions;
- handles failures, for example by retrying, routing around the problem or reporting clearly that part of the work could not be done.
When a subagent fails, the useful response carries context: what failed, what was attempted and any partial results. Silently dropping the failure, or abandoning the whole run because one branch failed, both leave the user with a worse outcome.
When not to use subagents
Subagents add coordination effort. For a short, simple task, one agent is usually easier and cheaper to reason about. Consider subagents when the work is clearly separable, when intermediate material is bulky, or when different parts of the job need different tools. Giving a single agent every tool for every task tends to make tool selection less reliable and mixes unrelated work in one context.
Where this shows up in Claude certification
Coordinator and subagent design, including context isolation, sits in the largest domain of the Claude Certified Architect exam. Study notes are on the agentic architecture domain page, and the scenario page shows how these ideas appear in exam-style situations. Certifications are issued by Anthropic. Agmo Studio, a Select Partner of the Anthropic Claude Partner Network, runs exam-preparation workshops, described on the Claude Certified Architect program page.