How Natera built a real-time voice scheduler with Amazon Bedrock AgentCore

# Problem and context Booking a phlebotomy appointment is complex for oncology patients who prefer at-home draws. Natera's existing workflow involved patients calling, authenticating, giving three preferred dates, and a human scheduler coordinating vendors and phlebotomists. That workflow required personal identifier authentication, SMS verification codes, and integrations with multiple third-party vendor systems.

Natera had a working system that used Twilio for telephony and a third-party AI provider running on Amazon ECS. The team saw opportunities to improve conversational accuracy, scalability, observability, and telephony flexibility. Pre-built patient agents like Amazon Connect Health didn't fit Natera's vendor coordination needs, so they rebuilt the agent on Amazon Bedrock AgentCore.

# Why AgentCore Natera chose AgentCore for three operational reasons: it's fully managed (removes container scaling burden), supports fast foundation models for context-aware intermediate replies, and provides built-in memory and observability. AgentCore's tracing captures which tools were called, step latencies, and session lifecycles, helping teams diagnose where delays occur—model inference, tools, or memory retrieval.

# Core architectural principles The implementation follows three core design patterns that suit real-time voice agents.

• Telephony streaming and model inference run on separate WebSocket connections with an orchestration layer between them.
• This separation makes it possible to replace the telephony provider or the model/runtime independently without redesigning the whole pipeline.

• When a backend tool or external vendor call takes time, the agent uses fast foundation models to produce intermediate, context-aware responses so callers perceive lower latency.
• This keeps conversational engagement high while the system completes longer-running operations.

• The agent supports mid-conversation authentication flows such as SMS verification codes and personal-identifier checks.
• Authentication is staged during the call so the agent can continue a natural conversation while meeting healthcare compliance requirements.

# Outcomes reported in validation

# Implementation notes and trade-offs

• Observability: AgentCore's traces capture each decision and tool call, enabling targeted debugging without guesswork.
• Flexibility: The dual-WebSocket approach protects investments in telephony or model upgrades.
• Engagement: Generating intermediate responses during slow backend calls reduces caller abandonment risk.
• Compliance: Progressive trust lets the system collect authenticated identifiers mid-call while maintaining conversation flow.

# When this approach fits Use this pattern when you need custom vendor coordination and telephony flexibility beyond pre-built patient engagement agents, or when you require fine-grained observability and managed scaling for production voice agents. The design is applicable to other real-time voice use cases that require external tool integration, mid-call authentication, and low perceived latency.

# Practical next steps for teams

• Evaluate whether your telephony provider can operate with a WebSocket bridge and whether your vendor integrations support asynchronous tool calls.
• Prototype intermediate response generation with a fast foundation model to measure perceived latency improvements.
• Map authentication touchpoints and design a progressive trust flow that fits your compliance needs.

# Bottom line Natera's migration to Amazon Bedrock AgentCore produced a managed runtime with detailed observability, flexible telephony integration, and conversational techniques that keep callers engaged during backend operations. The validation metrics cited—100% tool-calling accuracy across 500 simulations, sub-7-second perceived latency, and under $0.01 per call—illustrate the system's operational goals and performance in testing.