How to Use Claude Code Subagents for Your Business - Delegate Without Losing the Thread

What is a Claude Code subagent?

A Claude Code subagent is a separate Claude instance that your main session spawns to handle one specific task, with its own context window and its own set of tools. Instead of one long conversation trying to hold an entire project in its head, the main session writes a brief, hands off a slice of the work, and the subagent runs that slice on its own before reporting back. Think of it as delegating a task to a contractor with a written scope, not looping a second voice into the same call.

The reason this matters for a business rather than a codebase is that most real work is not one task, it is five or six smaller tasks that do not depend on each other: pull last month's numbers, draft the summary, check it against last quarter, format it for the deck. Running all of that in a single conversation means every step competes for the same context, and the further you get, the more the assistant is juggling. A subagent takes one of those steps off the pile entirely.

• The main session is the one that talks to you and decides what needs doing.
• Each subagent gets its own written brief describing exactly what it should produce.
• A subagent can be limited to a narrower set of tools than the main session has, so a research task cannot accidentally touch your files.
• It works independently and hands back a finished result, not a running commentary.

When does delegating to a subagent actually help?

It helps when a task is well-defined and does not depend on another task finishing first. Research a competitor, summarize a long document, check a list of numbers against a source, draft one section of a report - these are jobs a subagent can take away and come back with a finished answer for, while the main session keeps working on something else or waits on you.

It does not help when the task is really one continuous piece of judgment that keeps changing shape as you go. If you are still deciding what the deliverable even looks like while you work, keep that in the main conversation where you can redirect it sentence by sentence. Send subagents after the shape of the work is settled, not before.

Where do I set one up?

Claude Code's desktop app is the place to start. Inside a project, you add a subagent as its own file with a name, a short description of when to use it, and the brief itself, the same way you would write a one-page handoff for a contractor. The terminal is there for anyone who wants to edit that file directly, but you do not need it to get a subagent running for the first time.

[How to use Claude Code for your business](/blog/how-to-use-claude-code-for-your-business) covers getting the desktop app set up in the first place, and a subagent is one more file inside that same project once you are there.

1. Open the project in Claude Code's desktop app and find where subagents live for that project.
2. Give it a name that says what it does, not who it is for - 'contract-summary', not 'legal helper'.
3. Write the brief the way you would brief a contractor: the input, the exact output, and what it should never guess at.
4. Limit its tools to only what the task needs. A summarizing subagent does not need permission to send anything.

What is the Single-Brief Rule?

The Single-Brief Rule is our scoping rule for subagents: one subagent, one written brief, one visible output you check before it moves anywhere near a customer. If you cannot write the brief in a few sentences, the task is not ready to delegate yet, it is still something you need to think through yourself.

The failure mode when this rule gets skipped is quiet, not loud. A subagent given a vague brief still produces something, and it still looks finished. It just fills the gaps in the brief with its own guesses, and you find out which guesses were wrong only after the output already went somewhere.

• State the exact input the subagent will work from.
• State the exact output, down to the format if it matters.
• Name what it should flag rather than guess at, so unclear cases come back to you instead of getting decided quietly.
• Read the result yourself before it reaches a customer, a filing, or a public post. A subagent's output is a draft from a contractor, not a finished decision.

Can subagents work at the same time?

Yes, several can run in parallel on separate tasks, which is the main reason to use them instead of one long conversation. The limit is not how many can run, it is how many separate tasks you actually have that do not depend on each other or touch the same thing at the same time.

The trouble starts when two subagents are pointed at the same file, the same customer record, or the same output at once. Neither one knows what the other is doing, so you can end up with two different answers to the same question, or one subagent's changes overwriting the other's. Split by task, not by headcount - give each subagent a lane nothing else is working in.

How is a subagent different from a skill?

A skill is a procedure Claude reaches for automatically inside a normal conversation. A subagent is a separate worker the main session deliberately sends a task to and gets a result back from. They solve different problems, and most businesses that use both end up combining them: a subagent handling one leg of a project can reach for a skill the same way you would.

[How to use Claude Skills for your business](/blog/how-to-use-claude-skills-for-your-business) covers the skill half of that table in full, including the One-Job Rule for scoping one.

What goes wrong with subagents in practice?

The failures are the same shape as most delegation failures, just faster. None of them mean the feature is unreliable, they mean the brief or the boundary was missing.

• The brief was vague, so the subagent filled the gap with a guess nobody reviewed.
• Two subagents were pointed at the same record and produced two different answers to the same question.
• A subagent was given tools wider than the task needed, so a research job had write access it never should have had.
• Nobody read the output before it reached a customer, because it looked finished, and finished is not the same as correct.

[How to stop AI hallucinations in your business](/blog/how-to-stop-ai-hallucinations-in-your-business) covers the checking habits that catch a confident wrong answer before a client does, and every one of them applies to a subagent's output just as much as a single chat reply.

Delegate one task this week

Pick one task off your list that is genuinely self-contained, write the brief in a few sentences, and hand it to a subagent instead of doing it inside your main conversation. Read the result before it goes anywhere. One well-scoped subagent doing one job reliably is worth more than five vague ones you have stopped trusting.