Natera's voice scheduler: the hard parts were sockets, filler speech and when to ask for ID

The plumbing argument holds up if you look at what a phone number demands. An agent behind one has to keep two long-lived connections alive at the same time: the leg carrying audio to and from the caller, and the leg carrying that audio into the model runtime with tool calls coming back out. AWS names this the dual-WebSocket bridge ^[6], and it explains why the migration account is about WebSocket lifecycle and session state rather than prompting ^[5]. A runtime that recycles containers on its own schedule is awkward company for a socket that has to survive an entire conversation.

The second decision follows from the first. Vendor availability lookups and SMS verification take time the caller can hear ^[11]. According to AWS, the answer is to generate context-aware intermediate responses with fast foundation models on Bedrock while the slow work runs ^[12], which is why the headline figure is *perceived* latency under seven seconds ^[8] rather than end-to-end tool time. That is a filler-speech budget, and it only pays off if you know which step is slow. AgentCore's per-step traces record which tools were called, how long each took, and what the model decided ^[13].

The third decision is the one worth arguing about in a regulated queue. The workflow needs personal identifiers and an SMS code before anything patient-specific can move ^[11], and Natera authenticates part-way through the call instead of at the top, which AWS calls a progressive trust model ^[6]. It buys engagement, and it creates a window in which the agent is conversing with a caller nobody has verified yet. The scope question that follows is what the agent may volunteer, and what it writes down, before the code clears.

Then the validation numbers. Zero tool-calling failures in 500 simulated end-to-end calls ^[7] is a ceiling, not a floor. The rule of three puts the true failure rate at roughly 0.6% or below with 95% confidence ^[16], which at a few thousand calls a month is still tens of mis-fired tool calls. Simulations also do not cough, change their minds, or read a street name wrong.

The cost figure needs the same care. Under one US cent per completed call ^[9] buys the automated leg, and a completed call ends with authentication done, three candidate dates captured and a service location on file. A human scheduling team still coordinates with phlebotomists and vendors to confirm one of those slots ^[10]. Cost per booked appointment therefore carries labour the sub-cent number does not ^[17].

AWS is also unusually direct about when not to build this. Organisations on Epic or Cerner with standard booking flows are pointed at pre-built Amazon Connect Health patient engagement agents; Natera went custom because it needed vendor coordination and telephony control those do not cover ^[14]. The dividing line is integration surface, not model quality.