The short version

Key concept: Use subagents when context isolation, parallel execution, or a fresh perspective creates a real advantage.

Editorial analysis

The useful lesson is not a single prompt formula. It is a way to reduce ambiguity: state the outcome, supply only the context that changes the answer, demonstrate the pattern when format matters, and test whether the result remains reliable across several inputs.

A useful way to read this study is as a decision guide: identify the problem it solves, the conditions where it works, the tradeoffs it introduces, and the evidence you would need before relying on it.

Source context

Why this matters

Long Claude Code sessions accumulate files, tangents, intermediate thoughts, and assumptions. Anthropic describes subagents as isolated Claude instances with their own context windows. They complete a bounded task and return only the relevant result, helping the main conversation stay focused.

The important lesson is not “use more agents.” It is “create a separate context only when separation improves focus, speed, or objectivity.”

How it works

Think of the main Claude conversation as the project lead. It should retain the overall goal, key decisions, and final responsibility.

A subagent is a temporary specialist tab. It investigates one question, works with only the permissions and information it needs, and returns a concise handoff rather than every file or thought it encountered.

Good reasons to use one include:

  1. Research-heavy work: Understanding the problem requires reading many files or sources.
  2. Independent tasks: Several pieces of work have no dependencies and can run in parallel.
  3. Fresh review: An evaluator should not inherit the assumptions and blind spots of the implementation conversation.

Where it matters

Run separate bounded reviews for obligations, termination rights, privacy provisions, and unusual clauses. Each subagent returns a structured summary with citations; the main agent synthesizes the final view.

Finance

Use parallel subagents to investigate revenue movements, cost drivers, and working-capital changes. The main conversation combines the findings into a coherent management narrative.

Research

Delegate source discovery and fact verification separately from writing. This protects the final drafting context from large amounts of raw research material.

Workflow automation

Use one subagent to map a process, another to identify control risks, and a fresh reviewer to challenge the proposed automation before implementation.

Better implementation

Weak

Use some agents to research this project.

Problems:

  • No bounded scope
  • No explanation of which work is independent
  • No permission limits
  • No required output format
  • Likely duplicated effort

Strong

Use three read-only subagents in parallel. One should map the data sources, one should identify control risks, and one should find applicable reporting requirements. Return a one-page summary from each with citations. Do not edit files.

The strong version defines scope, parallelism, permissions, and the expected handoff.

When not to use subagents

  • The task is small and focused.
  • Each step depends heavily on the full output of the previous step.
  • Multiple agents would edit the same file.
  • The coordination overhead is greater than the likely benefit.
  • Too many similar specialists would make delegation confusing.

Implementation checklist

  • Does this work need a separate context?
  • Are the delegated tasks genuinely independent?
  • Would a fresh perspective reduce bias?
  • Is the subagent scope narrow and explicit?
  • Are permissions limited to what is required?
  • Is the expected return format defined?
  • Would completing it in the main conversation be simpler?

Try it in practice

Choose one workflow that often becomes a long AI conversation.

  1. Identify the main goal that should remain in the primary conversation.
  2. Mark any side investigations that do not require the full project context.
  3. Identify one decision that would benefit from an unbiased second opinion.
  4. Write one subagent brief using this template:
  • Task:
  • Why it should be separate:
  • Allowed access:
  • Required output:
  • The subagent must not: