1 / 14
How It Works
Manager splits task Sub-agent 1 fresh brain Sub-agent 2 fresh brain Sub-agent 3 fresh brain Sub-agent 4 fresh brain returns results Manager synthesizes
one conversation Quality fades over time → sub-agents Quality holds across all items
Usage cost
1× 2× 3× 4× Single thread Sub-agents 3 – 4× the usage
The independence test
Can each piece be done on its own?
Yes → Split it No → Keep it together
Split
Reviewing 30 contracts — each stands on its own
Keep
Drafting a report, then making slides from it
Split
Processing 40 invoices for vendor, date, and total
Keep
Writing a proposal, then building an exec summary from it
Split
Researching 5 competitors — each independent
Keep
Analyzing data, then writing a recommendation
Practice one
Tell each sub-agent exactly what you want back
Example prompt
Research Acme Corp, Beta Inc, and Gamma LLC. Use sub-agents to research each company in parallel. I'm evaluating them as potential partners for a mid-size consulting firm. For each one, find their pricing model, their biggest customers, and any major news from the last six months. Return a summary under 500 words per company.
Let the AI divide the work
Add to your prompt
Make sure each sub-agent covers a completely different area. No overlap between them. Together, they should cover the full picture.
Practice two
Keep the sub-agent count manageable
1 2 3 4 5 6 7+ sweet spot
Practice three
Only split independent work
independent — split Task A Task B Task C Task D dependent — keep together Analyze Write If you can draw an arrow → keep them together
Practice four
Is the time saved worth the extra usage?
worth it 40 receipts — saves over an hour 10 competitor reports at once skip it 5 docs — done in 15 minutes 3 short emails to draft in a row
The scenario
40 receipts & invoices .pdf .jpg .png .pdf Vendor Date Amount Category
The prompt
Tell the AI
I have 40 receipts and invoices in this folder. For each one, extract the vendor name, date, total amount, and expense category.

Use sub-agents to process these in parallel. Each sub-agent should return only those four fields — nothing else.

After all files are processed, combine everything into a single spreadsheet sorted by date.
Recap
An AI can split tasks into smaller tasks with sub-agents
Only split work that's independent
Be specific with each sub-agent about what to return and why
Keep to 3–4 sub-agents for best results
Check that time saved justifies the extra usage