Chapter 1 · Practice 3 of 4
Build one agent before splitting work across several
Why it matters
Multi-agent systems add handoffs, shared state and more places to fail. OpenAI's evaluation guide advises adopting multi-agent designs only when evals justify the added complexity, and adds handoff accuracy as something you must then measure.
Anthropic's prompting guide gives a useful rule for when sub-agents help: when tasks can run in parallel, need isolated context, or are independent workstreams that do not share state. For simple tasks, sequential steps or single-file edits, a single agent working directly is better.
Start with one agent and good tools. Split only when you can point to a measured problem, such as context overflowing or independent research that would run faster in parallel.
Do
- Measure handoff accuracy once you add a second agent
Avoid
- A team of named agents for a task one agent can do with three tools