The limit of one agent
Give a single AI pair a big task ("write a launch blog post for our new feature") and you will notice a characteristic failure: it does everything at once, at the same level of quality. The research is shallow. The draft is fine but generic. The critique is an afterthought because by then the thing is already written.
This is a decomposition problem. Humans solve it by changing hats: "first I'll research, then I'll draft, then I'll review." Subagents let the AI do the same thing.
What a subagent is
A subagent is a child task you spawn from inside a larger agent task, with its own narrowed scope, its own system prompt, and its own output contract.
Three shapes show up constantly:
1. The researcher
"Given this topic, find and summarize the 5 most relevant facts. Output a bulleted list with citations."
2. The writer
"Given these 5 facts and this target audience, write a 400-word post in this voice."
3. The critic
"Here is the draft. Find the three weakest paragraphs. Do not rewrite — only diagnose."
You compose them. The parent agent picks the topic, spawns a researcher, takes the output, spawns a writer, takes that output, spawns a critic, and then (importantly) you read the critique and decide what to do.
When subagents actually help
Subagents are a distinct win when:
- The task has clearly separable stages. Research, draft, critique. Design, implement, test. Fetch, transform, report.
- Each stage benefits from a narrowed system prompt. A researcher with a "just the facts, cite everything" prompt produces different output than a writer with a "make it engaging" prompt.
- Intermediate artifacts are worth saving. You want the research bullets on record even after the final draft ships.
They don't help — and often hurt — when:
- The task is small enough that spawning them is overhead.
- The stages aren't actually separable ("make this change well").
- You don't have a way to review the intermediate output.
A concrete example
Say you're reviewing a PR with 40 files changed. A single agent reads it, tells you "it looks fine," and moves on. A subagent pipeline instead:
- Researcher subagent: "List every behavioral change this PR introduces. One line each. Don't opine."
- Critic subagent: "For each behavioral change, rate risk (high/medium/low) and say what would break if this change is wrong."
- Summarizer subagent: "Given the above, produce a 5-bullet review summary."
You now have an artifact pipeline and a summary. Each step was good because its scope was narrow.
Practice
On your next non-trivial agent task, write out the pipeline before you kick it off:
plan:
- researcher: <what should it look up?>
- writer: <what should it produce?>
- critic: <what should it check?>
Then compare: did the 3-step pipeline produce better output than a single prompt? When yes, when no — write down your observation.
Coming up: Building with the Claude API — now you're going to put these skills to work wiring an AI into software you write.
Inspired by Anthropic's "Introduction to subagents".