Compose agents only where the separation helps.
1. Sequential and parallel work
Use sequential steps when results depend on earlier outputs. Parallelize independent read-only work while preserving clear aggregation and failure behavior.
2. Delegation boundaries
A specialist agent needs a narrow task, accepted input and expected output. Passing all history to every agent increases cost and can leak unrelated information.
3. Stopping conditions
Limit turns, tool calls, elapsed time and retries. Detect repeated actions and return a bounded failure rather than waiting indefinitely for the model to finish.
Worked scenario
A lookup agent repeatedly requests the same unavailable ticket. The coordinator stops after the allowed attempts and asks for corrected input.
Apply it
Add one specialist and compare it against the single-agent baseline for success, latency and call count.
Check your understanding
You can show a measurable reason for each additional agent. Explain the decision and show evidence from your implementation or design. If you cannot demonstrate it yet, revisit the relevant section before continuing.