Before you start
A signed-in NodeFox workspace. Use synthetic data and prepare outputs before enabling external writes.
01Measure the stages
Separate provider latency, external API latency, document processing and local logic. A Rust runtime does not remove network or model delays. Use observed runs rather than invented throughput numbers.
02Avoid redundant work
Retrieve shared evidence once when the data’s freshness and scope permit it. Keep prompts and intermediate payloads bounded. Validate before an expensive model call when invalid input can be rejected cheaply.
03Parallelize independent branches
Only run branches concurrently when their dependencies permit it. Respect provider rate limits and destination capacity. Keep required joins so a faster branch cannot advance a partial result.
04Recheck correctness
Compare outputs, exceptions and side effects after an optimization. Faster completion is not a success if the workflow loses evidence, skips review or creates duplicate actions.
If the result is wrong
Raising concurrency can increase rate-limit errors. Measure end-to-end completion and failure rate together.