Trace before optimizing
Instrument the request path and distinguish time spent waiting from time spent doing work. A single slow dependency can dominate the experience even when model generation is fast.
Find independent work
Parallel execution helps when tasks do not depend on each other. Preserve ordering where one result supplies another step, and bound concurrency to protect shared dependencies.
A harness makes context, tools, execution, and review part of one observable system.
Use caching deliberately
Reuse stable work when its inputs and permissions remain valid. Define invalidation behavior and verify that cached results still fit the request.
Keep quality in the evaluation
A faster response is not an improvement if it misses essential evidence. Evaluate latency together with correctness and reliability using representative workloads.