Amazon outlines three patterns for building low‑latency, scalable voice agents

A new technical post presents three concrete architectural patterns to build low‑latency, scalable voice agents using Amazon Nova Sonic, Bedrock AgentCore Runtime (with AgentCore Gateway) and the open‑source Strands BidiAgent. The write‑up targets teams wrestling with real‑time audio problems — high latency, coordination across multiple agents, and the need for session‑level isolation — and shows how those components can be integrated to reduce latency and simplify operational behavior.

The post divides the solution into three building blocks. Nova Sonic is described as a speech‑to‑speech foundation model designed for real‑time, natural conversational flow. Bedrock AgentCore Runtime provides a serverless hosting environment with bidirectional WebSocket streaming protected by SigV4 authentication, microVM‑level session isolation, an AgentCore Gateway for hosting tools via the Model Context Protocol (MCP), persistent session memory, and voice‑specific telemetry such as time‑to‑first‑audio. The Strands BidiAgent class handles bidirectional streams, routes tool calls, and manages session lifecycle.

Authors present three integration patterns. In the tool‑driven pattern models call external functions (tools) directly. In the agent‑as‑tool pattern sub‑agents are exposed as tools so higher‑level models can invoke them. Session segmentation isolates prompts, memory, and permissions across boundaries. The post argues that decomposing systems into smaller, specialized components improves reusability and security, while warning that increased decomposition can add complexity and potential latency trade‑offs that builders must weigh.

For latency‑sensitive interactions the post highlights AgentCore Gateway plus MCP as a low‑overhead path: tools are hosted as discrete MCP endpoints and invoked directly by Nova Sonic during a conversation, avoiding a separate reasoning or orchestration hop. AgentCore Gateway runs managed MCP servers; agents connect using Gateway ARNs. Bedrock AgentCore Runtime also contributes scaling, billing, and telemetry features that make the operational behavior of voice workloads observable and easier to manage.

A concrete example illustrates configuration and the call flow. The example shows a BidiNovaSonicModel instance configured with model_id="amazon.nova-2-sonic-v1:0" and an mcp_gateway_arn list of Gateway ARNs. When a user asks “What’s my account balance?”, Nova Sonic extracts intent, selects the get_account_balance tool from available MCP endpoints, invokes it with parameters, and speaks the returned result — demonstrating a direct model‑to‑tool path in practice.

Practical rollout notes cover prerequisites and recommended practices: install Python and dependencies such as strands‑agents and boto3, ensure IAM permissions for the services used, and consult the referenced GitHub repo for full examples. The post recommends session segmentation to reduce noisy‑neighbor effects, using sub‑agents‑as‑tools for reusable logic, and relying on AgentCore Runtime’s session isolation and telemetry to keep voice workloads stable and observable.