Designing Roles

Good agent team design starts with role definition. Each agent needs a clear identity: what it does, what it delivers, and what it does not do. Vague roles produce vague output. "Research agent" is not a role — "Research Agent: searches the web for competitor pricing data and returns a structured JSON table with company, product, price, and source URL" is a role.

The specialist principle is the foundation of role design: each agent does one thing well. A research agent researches. A drafting agent writes. A review agent critiques. When you let one agent handle multiple responsibilities, you recreate the single-agent problem — a crowded context window making compromises. Splitting responsibilities across agents means each one gets a clean context dedicated entirely to its specialty.

1. 1. List the Subtasks Write down every discrete step required to complete the full objective. Group related steps together. Each group becomes a potential agent role.
2. 2. Define Deliverables For each role, specify exactly what the agent must return. A research agent returns structured data. A draft agent returns prose. A review agent returns a score and specific feedback. Ambiguous deliverables produce ambiguous results.
3. 3. Set Boundaries State what each agent should NOT do. The research agent should not write the final report. The draft agent should not second-guess the research. Boundaries prevent scope creep and keep each agent focused.
4. 4. Name the Roles Give each agent a clear, descriptive name: Researcher, Drafter, Critic, Formatter. Names communicate expectations to the model and make your prompts clearer to read.