McCloskey.ai · ResourcesAll resources
← Resources Delegation at scale

Subagents

Delegate to Claude the way you already delegate to people. What fills up, when to hand work to a helper, and the brief that comes back as finished work.

MEDIA SLOT · hero-subagents-trailer.mp4 (fallback hero-subagents.png)
a 25-30s trailer: the desk fills, a helper takes the load and returns a clean summary, ending in an ambient loop

Based on the Introduction to Subagents course by the Anthropic team.

A subagent is Delegation, the first D of the framework, at scale: you hand the digging to a helper and keep the judgement calls. Meet the framework first on AI Fluency →

The whole idea

The desk, the helper, the rule, the trade-off

Four beats. A problem you already feel, the fix you already know, one habit, and the honest cost.

Context window

the problem
  • One desk holds the whole conversation.
  • Fill it and things fall off.

Subagents

the fix
  • A focused helper with its own desk.
  • Brief it like a hire, not a wish.

The decision rule

the habit
  • Does the middle matter? Keep it.
  • Only the answer? Delegate it.

The trade-off

the honesty
  • Only the summary comes back.
  • Ask for sources you can check.

Context window

Know what fills up

MEDIA SLOT · subagents-context-window.png
a single desk filling with messages, files and searches until items slide off the edge, one item terracotta

Every message, file and search lands in one place: the context window. That space is finite.

  • What sits on the desk? Every message, file and search you have sent.
  • What happens when it fills? Claude loses the thread, and your opening brief slides away.
  • When does it bite hardest? On long jobs, where the space is tightest.

One conversation is one desk. Pile everything on it and things fall off.

Subagents

Brief a helper like a hire

MEDIA SLOT · subagents-brief.png
a focused helper reading a one-page brief at its own desk and handing back a clean summary, one terracotta line on the brief

A subagent is a focused helper with its own context window. It does the messy work elsewhere and returns a summary. You already know this job: brief it like a hire, not like a wish. A good brief carries four things.

  • What is the job? A plain description of the work.
  • What comes back, and in what shape? Name the deliverable up front.
  • What should it flag rather than hide? Blockers, with no silent workarounds.
  • What can it touch? Only the access the job actually needs.

The difference it makes: "Look into our competitors" becomes "Research our three closest competitors. Return a one-page summary, five sourced findings, and any blockers you hit." You meet subagents in Cowork's parallel workstreams, Research on claude.ai, and Claude Code, the same idea everywhere.

The decision rule

Sort every task in one question

MEDIA SLOT · decision-rule-showcase.png
a single fork in the path: one way is delegate, the other is keep, one terracotta arrow

One question sorts every task: does the intermediate work matter?

  • Do you want the answer, not the journey? Delegate it: research, exploration, a fresh pair of eyes.
  • Does each step depend on the last? Keep it in the main thread.
  • Do you need to see and react as it goes? Keep it: the working is the point.

In practice: delegate the digging, keep the judgement calls.

The honest part

Where delegation goes wrong

Three moves that feel clever and cost you, straight from Anthropic's own guidance, and the trade-off underneath them all.

  • Expert claims: "You are an expert" adds nothing. Claude already has it.
  • Sequential pipelines: handoffs lose information. Keep dependent steps together.
  • Test runners: "Done, no issues" hides what you needed to see.

Only the summary comes back, the journey is gone. Ask for sources you can check, add obstacle reporting to every brief, verify what matters. A good brief is how you audit work you never see.

Put it to use

Try it on something real

Write one delegation brief and run it. Pick a real question: a competitor, a market you're weighing, a prospect, a role you're hiring for. No question in mind? Scan ten organisations you could partner with next year. Run it in Cowork, or on claude.ai with Research switched on.

The delegation brief
I'm the [your role] at [your company]. Research [three competitors / the market for X / this prospect] and bring me back a finished briefing I could forward without edits.

The deliverable, in exactly this shape:
1. Summary: three sentences, plain English.
2. Key findings: five bullets, each with a source I can check.
3. What surprised you: two bullets.
4. Obstacles: anything you couldn't find, or had to work around.

Skip the play-by-play. I only want the briefing.

Notice what you wrote: a description of the job, an output format, obstacle reporting. The fourth element, limited tool access, is set by which folders and connectors you've given Claude. Three of the four in one prompt.

Check yourself · two minutes

Did it stick?

Flip the cards, then match the pairs, while your brief runs.

An hour in, Claude forgets the brief you opened with.Which idea? Tap to flip
Context windowFinite space. Full means forgetful.
Claude ran forty searches to answer one question. You never wanted to see them.Which idea? Tap to flip
The decision ruleIntermediate work doesn't matter? Delegate it.
You chained three helpers, each feeding the next. Half the detail is missing.Which idea? Tap to flip
Sequential pipelinesHandoffs lose information. Keep dependent steps in one thread.
The summary is clean but you can't see the working.Which idea? Tap to flip
The trade-offClean context bought with lost visibility.

Click a term, then its match. The pair locks in when you get it.

Downloads

Extra resources

Free to download, no sign-up.

The Delegation Brief

A one-page template: job spec, deliverable, blockers. Paste into Cowork.

mccloskey-subagents-delegation-brief.pdf
Download ↓

Delegate or Keep It Close

The decision rule and the three traps, on one card.

mccloskey-subagents-delegate-or-keep.pdf
Download ↓
What you learned

Write one brief with all four elements and run it this week. The habit compounds: delegate the digging, keep the judgement calls.

Where next

Subagents are the top rung of the Leverage Ladder → explained: a whole team of helpers, and the when and how of using them. The ladder shows every step below it, from a single good prompt upward. To see a team of them running live, visit Mission Control → New to the framework? Start with AI Fluency →