Subagents: delegating a search

This lesson is still being tested: it improves with learners' feedback. If something is unclear, it is the lesson that needs fixing, not you.

By the end of this lesson, you will be able to

say what a subagent is, and when it is useful;
delegate a search to a subagent, and check its summary;
create your own subagent, read-only;
choose between a subagent, a skill and the main conversation.

Why it matters to you

When Claude searches through many files, your conversation fills up with results that you will never read again. It gets long, slow, and Claude follows what matters less well. A subagent does the search separately, then gives you back only the essentials. It is a delegation: useful, as long as you check the result.

Before you start


You need	Claude Code (app or terminal), in Manual mode, and your mon-espace-ia folder with a few notes or documents
Cost	The agent subscription: Claude Pro at minimum, €18 per month excl. VAT, price shown in France on 24/10/2026. A subagent uses your usage limit just like your conversation
Duration	35 minutes
Reminder	The subagent reads your files like Claude does: nothing from the “never” list in your folder

Understand it in 3 minutes

A subagent is a specialized assistant that Claude launches for a task, according to the documentation (24/10/2026):

it works in its own context and only returns the summary;
its requests count toward the same usage limits as your conversation;
Claude Code already has some, for example Explore, which searches and reads, without being able to write or edit;
you can create your own: a text file, with a name, a description and the tools it is allowed to use.

Subagent, skill or conversation?

Choose	When
The main conversation	An ongoing exchange, a small fix, a task where each step depends on the previous one
A skill (lesson 2)	A way of doing things that you repeat, applied in your conversation
A subagent	A separate task, which would produce a lot of useless output, and which can be summarized

Delegating does not excuse you from checking. A summary cites files or passages: open at least one.

Do it together

1Step 1: Delegate a search

In mon-espace-ia, in Manual mode, write:

Use a subagent to search my folder for notes and documents that talk about budget. Give me only a summary of 5 lines at most, with the file names.

In our test, Claude launched the Explore subagent, which searched read-only, then returned a summary of three files, specifying which ones did not mention it.

2Step 2: Check the summary

Open one of the cited files: does it really talk about budget, as the summary says? If a file is missing, ask: “Why is [file] not in your list?”

3Step 3: Create your subagent, read-only

Create a project subagent named chercheur-de-notes, in .claude/agents/chercheur-de-notes.md. Description: "Searches the notes and documents of this folder, and returns a short summary with the file names. To be used to find information in my notes." Tools: only reading and file search (Read, Grep, Glob). It does not edit anything. Show me the file before creating it.

Read the header: the name, the description, the list of tools. The .claude folder is protected: Claude Code asks for your permission, in the app and in the terminal alike. Allow it if it looks right.

4Step 4: Use it

Open a new session (a subagent created during a session may only be seen at the next session), then:

Use the chercheur-de-notes subagent to find what I decided about [a topic from your journal].

Check the summary against your journal.

Your turn

Delegate a real search in your folder, and check at least one cited file.
If you often do this kind of search, keep your chercheur-de-notes subagent; otherwise, delete it (lesson 7).
Write in your journal what you delegated, and what the check showed.

Check your work

It worked if:

you got a summary from a subagent, and checked at least one file it cites;
your chercheur-de-notes subagent exists, with only read tools, or you decided not to keep it;
you can say, for a given task, whether the conversation, a skill or a subagent is better.

Common mistakes:

delegating a small task: it takes longer, and it uses up your usage limit;
giving a subagent write rights “just in case”;
taking its summary for proof;
multiplying subagents: each description weighs on the context of every session.

Keep it and pick it up later

Your project subagents are files in .claude/agents/: they are part of your folder, and so of your copies.
At the monthly upkeep, remove the ones you no longer use.

Going further

optional

You can also call a subagent with @ followed by its name, to be sure it is the one working.
The official documentation, in English: code.claude.com/docs/en/sub-agents (creation, personal or project scope, choice of tools and model).

New words

4 words explained

Word	What it means
Subagent	A specialized assistant that Claude launches for a separate task, and that returns a summary
Context	What Claude has in front of it during the session; a subagent has its own
Explore	The subagent provided with Claude Code to search and read, without writing
Delegate	Hand over a task, then check what comes back

The delegation in step 1 was actually run with Claude Code 2.1.310 in a test folder (test report, T40). The creation of the subagent in step 3 was not run.