Choose the smallest useful agent workflow

List the inputs, expected output and completion checks. Separate predictable steps from decisions that need interpretation.

Use ordinary code for repeatable transformations and validation. Reserve model-led work for tasks where the next step depends on context.

A bounded agent with a few well-defined tools is easier to inspect than a team with overlapping responsibilities. Give each tool a clear input, result and failure state.

Test real tasks before introducing coordination. More agents need a specific benefit, such as independent review or work that can run separately.

Use a fixed sequence when the order is known. Use independent tasks when their outputs can be checked and combined. Keep a single owner for the final result.

Define what happens if one task fails. A workflow should report partial completion without presenting it as a finished outcome.

Name the actions that need approval and the evidence required to close the task. Put limits on retries, time and resource use.

Review the workflow against quality, latency and operating cost. Keep the extra coordination only if it improves those outcomes.