How to use sub-agents in Claude for parallel, independent tasks

Executive overview

As a single AI conversation grows, its context fills up and instruction-following degrades. Sub-agents solve this by having a parent AI split a large job into independent subtasks, each handled by a separate AI with its own memory running in parallel.

The trade-off is usage: sub-agents can multiply consumption 3–5x, so they only make sense when the task is truly parallelisable and the time saved justifies the cost.

Core insight: sub-agents are only worth it when the subtasks are fully independent — if any output feeds into the next step, keep it in one thread.

The independence test

• Ask: can each piece be completed without knowing what the others are doing?
• If yes, sub-agents are viable; if outputs are connected, keep them in a single thread
• Split examples: reviewing batches of contracts, extracting data from invoices, researching multiple competitors
• Keep examples: drafting a report then building slides from it, writing a proposal then summarising it, analysing data then writing recommendations

Four practices for effective use

1. Be specific about what each sub-agent returns — define the exact fields or format you want back; include your intent so the AI understands why
2. Keep sub-agent count to 3–4 — more risks saturating the parent agent and burns usage fast; 10–15 is possible but rarely justified
3. Apply the independence test first — if steps have arrows between them (output of one feeds the next), that segment cannot be parallelised
4. Make sure the juice is worth the squeeze — processing 40 receipts or researching 10 competitors justifies the usage; writing a few emails or summarising five short documents does not

Prompting the parent agent

• State the task and explicitly say "use sub-agents" and "process in parallel" — the parent needs this instruction
• Give intent: tell the AI why you're doing the task, not just what
• Specify the output per sub-agent (e.g. "return only vendor name, date, amount, category — nothing else")
• Instruct the parent to consolidate all sub-agent outputs into a final deliverable
• If unsure how to divide the work, tell the parent AI to ensure each sub-agent covers a distinct area with no overlap and full coverage together