Course contents · How AI agents actually work

Lesson 10 of 13

Subagents: when an agent delegates.

Why serious agent systems hand jobs to other agents, and how to write the brief, shape what comes back and limit what each one can do.

Joe LeungJoe LeungCTO and co-founder, LaunchLemonade7 minReading time

When you ask a colleague to look into something, you give them a brief, not a transcript of every meeting you've ever attended. And when they report back, you want the findings, not their browser history.

What a subagent is

A subagent is just an agent started by another agent to do one job: research this topic, review this document, summarise this pile of PDFs. It gets a brief, does the work and reports back.

Many serious agent systems now work this way, with a main agent acting as the manager and handing pieces of work to others.

Why bother? Context, mostly

Everything a model sees competes for its attention. A long-running agent piles up baggage: every tool result, every wrong turn, every step along the way. If it does a big research task in the middle of a job, hundreds of pages flow through its context, and most of them were only ever needed to produce one summary.

Delegating a research task5 steps
  1. Main agent: "Research the top three competitors' pricing. Report back in one page."
  2. The subagent starts fresh, with just that brief
  3. It searches, reads and compares, using up its own context to do it
  4. It returns the one-page answer
  5. The main agent gets the page, not the searches behind it

Three delegation questions

What goes in the brief?
Too little and the subagent guesses. Too much and you've defeated the point.
What comes back?
A structured summary the main agent can act on. Not a raw dump.
What can it do?
Only what the job needs. A research subagent probably doesn't need the email tool.

The last question is the permissions lesson again. Scoping tools applies to subagents too, and each one is a chance to give less access, not more.

Parallel work and a smaller blast radius

Subagents also allow parallel work: three research tasks at once instead of one after another.

And they contain mistakes. A subagent that goes down a rabbit hole costs you one bad report, which is a small blast radius. The main agent can spot it, ignore it or ask again, without the whole job going off course.

One agent is a worker. Subagents make it a team, with the main agent as the manager.

Running example

The running job

Chasing last month's unpaid invoices, one lesson at a time.

Priya chases about a dozen clients a month, and one agent copes fine with that. But suppose the studio grew and had three hundred to check. Rather than one agent reading every invoice, email thread and note in a single long session, the main agent could hand each client to a subagent with a short brief.

The brief each subagent getsExample
Client: one client's name and their overdue invoice number.
Job: check whether it's paid, how overdue it is, and whether the client has raised any problem with it.
Return: amount, days overdue, any dispute (yes or no, with one line why), suggested tone (friendly, firm or hold off).
Tools: read the accounts and that client's emails. No sending.

The main agent gets one short summary per client instead of three hundred inboxes. It drafts the reminders from those, skips anyone who has paid or is disputing, and passes the drafts to Priya. None of the subagents can send email, so a confused one can only produce a bad summary, never a bad message to a client.

Practice

Try it yourself

About five minutes, in any AI chat app

Write a brief, then check what comes back

  1. Pick a small research task, such as comparing three tools you're considering. Write a brief for it in any AI chat app, saying exactly what you want returned: for example a table with price, one strength and one weakness for each.

  2. Start a fresh chat and paste in only the brief. Let it do the work.

  3. Compare what comes back with what you asked for. Then try again with a vaguer brief, such as "look into these tools for me".

What to look for

The tight brief should give you something short you can act on straight away. The vague one usually comes back longer and harder to use. That difference is the whole skill of delegating to a subagent.

Common mistakes

  • Passing the subagent the whole conversation.

    Give it a focused brief. Its fresh, clean context is the reason to use one.

  • Accepting whatever comes back.

    Say exactly what shape the answer should take, so the main agent can use it without wading through it.

  • Giving every subagent the same tools as the main agent.

    Scope tools to the job. A research subagent rarely needs to send anything.

Quick check

3 questions to see what stuck. Nothing is scored or saved.

  1. 1. What's the main reason to use a subagent?

  2. 2. What should a subagent return?

  3. 3. A research subagent goes down a rabbit hole. What's the damage?

Remember

Delegate like a good manager: a clear brief, a defined answer and only the tools the job needs.

Sources and further reading

  1. 1.The LinkedIn post this lesson grew out of: Subagents. Why would an agent need to hire another agent?
  2. 2.Anthropic: How we built our multi-agent research system

Finished this lesson?

0 of 13 done in this course. Progress stays in this browser, no account needed.