The latency budget for a voice agent
A caller notices silence at about 800ms. Transcription, inference and speech all have to fit inside that, and most of the budget is already spent before the model runs.
Engineering
3 posts.
Making agent actions idempotent
An agent that retries a tool call can book the same load twice. The fix belongs in the action, not the prompt.
Testing a system that talks
Conversations have no single correct output, so assertions have to move from the transcript to the outcome.