Natera Moves Scheduling Voice Agent to Bedrock AgentCore; AWS Reports Sub-7-Second Latency

Natera has migrated an automated voice agent for mobile phlebotomy scheduling from Amazon ECS containers to Amazon Bedrock AgentCore. AWS says the resulting design delivered sub-seven-second perceived latency at under $0.01 per completed call in validation, tying phone calls, model inference, authentication and vendor systems into one workflow.

The agent is designed to let patients arrange at-home blood draws by phone. Its scheduling path includes patient authentication and SMS codes, queries to third-party systems for available appointment windows, and fallback handling when a case becomes more complex.

A bridge between the call and the model

The core pattern keeps two WebSocket connections open: one between Twilio and the agent for phone audio, and another between the agent and a real-time voice-processing API. The AgentCore-hosted agent sits between them, forwarding audio while intercepting tool requests, running business logic and returning results to the conversation.

That separation is intended to let Natera change either the telephony or model-processing side without rebuilding the entire design. It also puts coordination in the middle layer, which must manage concurrent connections and keep WebSocket sessions alive for the duration of a call.

Designing around the wait

Natera’s latency strategy does not require every backend call to finish instantly. When the agent begins a slower tool call, it can generate a short, context-specific acknowledgment in parallel, so a caller hears a response rather than silence while availability or identity checks run.

AWS says the team measured per-tool latency over a two-week period, used each tool’s median response time, and set a filler trigger one second before that median. The agent suppresses the filler when a call finishes before its trigger, avoiding an extra spoken interruption for quick operations.

Authentication without restarting the conversation

The agent begins with a lower-trust session keyed to a SHA-256 hash of the caller’s phone number. After verification against Natera’s identity service, it creates a session keyed to the verified patient ID, moves the earlier conversation history into that authenticated session, and excludes the original session from later retrievals.

Managed runtime, harder coordination

The migration removed container-specific health checks, scaling policies and deployment manifests from the agent code. Natera also replaced container-local conversation state, which could disappear after restarts, with AgentCore memory for durable persistence across sessions.

AWS identifies the tradeoff plainly: the pattern increases orchestration complexity through concurrent WebSockets, parallel filler generation and progressive memory sessions. For simpler single-turn or text-only agents, AWS says direct integration can reduce that overhead. The reported performance figures come from simulations, leaving real-world outcomes across complex calls as the next test.