Sequential or parallel? Draw the flow.

Key takeaways

• 1Most multi-agent systems use the wrong execution flow: sequential when parallel would be faster, or parallel when correctness needs sequence. The fix is to draw the dependency graph before you write the orchestration.
• 2Use sequential orchestration when each agent needs the previous agent's output. A research, writer, reviewer chain is the classic case: the writer cannot start before research finishes, and the reviewer cannot start before writing finishes.
• 3Sequential flow costs roughly 3x latency because the agents run one after another. That latency is the price of correctness, and you only pay it when there is a real dependency between the steps.
• 4Use parallel orchestration when agents work on independent slices of the same task, like three product descriptions or five test cases. No agent depends on another's output, so they can all run at once.
• 5Parallel execution can save up to 70% latency. Independence is the speedup: when there is no dependency, running agents concurrently is almost free wall-clock time.
• 6The test before you code: if agents depend on each other, go sequential; if there is no dependency, go parallel. Draw the flow first, then write the code.