Skip to main content

11. Sub-Agents

Note: The video covers material not in the guide below — please watch in full.

Action Step

Complete this before moving on.

Follow along with the video and spawn sub-agents yourself. Tell Claude to research three competitors in a space you choose, have each sub-agent save its research into a separate markdown file in a folder you create in your repo, and then have the master agent collate everything into one comprehensive summary. Don't just stick to the example — get creative and try your own prompt for sub-agents too.

Comment in Slack

Post your answer in your onboarding channel.

What was your biggest takeaway(s) from this training?

Training Guide

You just learned that one agent has a 200,000 token budget — and you learned how to manage that budget with handoffs, learning docs, and chain-linking.

But what if the task is so big that even chain-linking feels slow? What if you need to research 10 competitors, enrich 68 documents, or generate content across dozens of files?

You don't do it one at a time. You clone yourself.

(Let's talk about how the AI builds a team)

The Manager and the Team​

Your main Claude session — the one you're typing into — is the master agent. It's the manager. And the manager can spin up copies of itself called sub-agents.

Each sub-agent gets its own fresh 200,000 token window. They work at the same time, independently, and report back to the master when they're done.

Think of it like this: you're a manager with a team of five people. You don't do all five tasks yourself — you hand each person a task, they go work on it, and they come back with the result. That's exactly what's happening here, except the "people" are copies of the AI.

The key thing to understand: sub-agents only know what the master tells them. They don't talk to each other. They don't share context with each other. The master gives each one a specific job and the context it needs to do that job — nothing more.

(So when does this actually make sense to use?)

When to Use Sub-Agents vs a Single Agent​

Sub-agents are powerful, but they're not always the right move. Here's the simple rule:

Use sub-agents when the tasks are independent. If each task can be done without knowing the result of the other tasks, sub-agents are perfect. Research five competitors separately. Enrich 50 documents that don't depend on each other. Write 10 chapters that each stand alone.

Use a single agent when the tasks are sequential. If step 2 needs the output from step 1, sub-agents can't help — you need one agent working through the steps in order. That's where chain-linking comes in.

A quick way to test: can you hand this task to five different people who don't talk to each other? If yes — sub-agents. If no — single agent with chain-linking.

(Now let me show you what this looks like at scale)

The 68-Playbooks Example​

Here's a real project that shows why sub-agents matter.

We had 68 playbooks that each needed to be enriched with additional research — market data, competitor info, updated stats. Each playbook was independent. None of them depended on each other.

Here's how it worked:

  • The master agent orchestrated everything
  • It spawned 5 sub-agents at a time, each handling one playbook
  • Each sub-agent had its own fresh 200,000 tokens to work with
  • When a batch of 5 finished, the master spawned the next batch
  • 12 batches total to get through all 68

The numbers: 13.8 million tokens used across all sub-agents. The whole thing ran for four and a half hours — autonomously, without any human input after the initial prompt.

A task that would've taken weeks of manual work, done overnight. That's the scale unlock.

(Let's try it yourself)

Guided Demo: Spin Up Sub-Agents​

Tell Claude something like this:

"I need you to research three competitors in the [your industry] space: [Company A], [Company B], and [Company C]. Spawn sub-agents to handle each one in parallel. For each competitor, create a summary with key findings — what they do, who they serve, how they're different. Save the results of each subagnt into a different document"

Now watch what happens:

  • Look at the sidebar. You'll see Task indicators appear — each one is a sub-agent working independently
  • Click on a Task to see what the master agent told the sub-agent to do
  • Wait for them to finish — they'll each report back to the master
  • The master merges the results into your document

That's it. You just ran three research projects simultaneously instead of doing them one at a time.

(Now step back and see the full picture)

The Combination Unlock​

Think about what you now have:

  • Token management — you understand the budget and how the AI's memory works
  • Context window management — you know how to save your work before compacting, create handoff documents, and chain-link across conversations
  • Sub-agents — you can spin up a team that works in parallel, each with their own fresh budget

Combine chain-linking with sub-agents and you have the architecture to run massive projects autonomously. A master agent that chains across sessions for the sequential stuff, and spins up sub-agents for the parallel stuff.

One conversation can do a lot. Chain-linked conversations can do more. Add sub-agents, and you can scale to almost anything.

(That's the full toolkit for how these agents work under the hood. Now let's come back to where we started — and land the big picture)