Natera Puts a Bedrock AgentCore Voice Scheduling Agent Into Production

A production voice agent built around AgentCore

AWS has published a detailed production case study showing how diagnostics company Natera rebuilt its mobile-phlebotomy scheduling voice agent on Amazon Bedrock AgentCore.

The system is now in production for inbound appointment scheduling. It combines telephony, real-time voice processing, agent orchestration, patient authentication, retrieval, memory, and backend scheduling tools.

This is not a new AgentCore product launch. It is a fresh production architecture and deployment report, published on August 26, 2026, with unusually specific validation, latency, cost, and early-production metrics.

Dual-WebSocket architecture

Natera’s agent uses a dual-WebSocket bridge.

One connection carries phone audio through Twilio. A second connects the orchestration layer to a real-time voice-processing API. The AgentCore-hosted agent sits between them, intercepting tool calls, executing business logic, and returning results to the conversation.

AWS says this separation makes either side replaceable without redesigning the whole system. For example, the telephony layer or voice-model provider can change while the orchestration and scheduling workflow remains stable.

Natera migrated the orchestration workload from Amazon ECS to Bedrock AgentCore Runtime, which provides isolated microVM-based execution. Session state moved out of container-local memory into AgentCore Memory, allowing state to survive restarts and persist across sessions.

Designing for perceived latency

The system’s most practical design idea is event-driven latency masking.

Backend scheduling and authentication calls can take several seconds. Instead of leaving callers in silence, the agent measures expected tool latency and can generate a short contextual acknowledgment in parallel when a call exceeds a calibrated threshold.

Natera derived those thresholds from per-tool latency traces. AWS says the team used AgentCore trace export and CloudWatch logs over a representative two-week period to estimate each tool’s median response time, then triggered filler generation before users would perceive an awkward pause.

The technique does not make the backend API itself faster. It reduces perceived latency by making the interaction behave more like a human call-center agent who acknowledges that a check is in progress.

Progressive trust and memory

The architecture also separates low-trust conversation state from authenticated patient state.

The agent begins with a session keyed to a SHA-256 hash of the caller’s phone number. That allows general interaction without exposing protected patient information.

After full identity verification, the system creates an authenticated session keyed to the verified patient ID. Earlier conversation history is transferred into the authenticated session so the caller does not need to repeat information.

Sensitive tools are not exposed before verification, and post-verification tool access is scoped to the authenticated patient session.

AWS describes this as a progressive trust model: conversational access expands as identity confidence increases rather than forcing every caller through a hard authentication gate before any useful interaction can begin.

Validation results

AWS reports that Natera spent four months iterating on the system and ran 500 end-to-end call simulations.

During that validation set, the orchestrator achieved 100% tool-calling accuracy, and AWS says parameter extraction was correct for all simulated tool invocations.

Median perceived latency measured 6.8 seconds end to end, including voice processing, with the Bedrock portion at 6.2 seconds. AWS also reports the cost stayed below $0.01 per completed call, including model inference, tool execution, and memory operations.

Those are results from Natera and AWS’s implementation and validation process, not independent benchmarks for all AgentCore deployments. Workload shape, external APIs, telephony, model choice, region, and compliance architecture can materially change latency and cost.

Early production data

The post includes a four-week early-production window.

AWS says the AgentCore-based system handled 4,744 calls, 5.5% more than the prior system. The share of calls ending in under 30 seconds fell from 22% to 12%, which the team interprets as fewer premature abandonments.

Verification completion moved from 64% to 66%, while average duration for resolved interactions increased from 79 seconds to 101 seconds. NPS response rates increased from 1.09% to 1.60%, while overall promoter and detractor shares remained broadly stable.

The system is now being expanded toward omnichannel inbound and outbound support.

Security and healthcare boundaries

The design uses HIPAA-eligible AWS services under a Business Associate Agreement for protected health information handled inside that boundary.

Amazon Bedrock Guardrails are used for conversational controls, and access to personal-information tools is gated behind identity verification. The agent is designed not to provide medical advice, diagnosis, or result interpretation beyond approved tool outputs.

AWS also notes that the voice-processing layer operates under separate data-processing arrangements and that Natera is exploring a more fully managed AWS path to keep audio processing inside the BAA boundary. That distinction matters: using a HIPAA-eligible agent runtime does not automatically make every external component of a voice stack part of the same compliance boundary.

Why the architecture is useful beyond healthcare

The case study is valuable because it exposes practical agent-engineering problems that benchmark scores do not capture:

• long-lived streaming connections inside managed runtimes;
• session-state migration from containers to managed memory;
• tool-specific latency instrumentation;
• progressive authentication;
• real-time fallback and escalation;
• observability at each agent loop step.

For teams building customer-facing voice agents, those system-level decisions can matter as much as the choice of foundation model.

Availability

Natera’s scheduling agent is in production. Amazon Bedrock AgentCore is an existing AWS service; the August 26 publication documents this customer deployment and architecture rather than announcing a new AgentCore release.