Résumé IA
Amazon Web Services et Pipecat ont publié un guide détaillé sur le déploiement d'agents vocaux intelligents en production, s'appuyant sur la nouvelle infrastructure Amazon Bedrock AgentCore Runtime. La solution combine Pipecat, un framework open source spécialisé dans les pipelines audio temps réel, avec l'environnement serverless d'AWS pour permettre des conversations vocales naturelles sur le web, le mobile et la téléphonie. L'architecture prend en charge trois protocoles de transport réseau : WebSockets, WebRTC et l'intégration téléphonique classique. Chaque session de conversation tourne dans des microVMs isolées, avec une capacité de session continue allant jusqu'à 8 heures, et une mise à l'échelle automatique face aux pics de trafic. Le runtime impose l'usage de conteneurs ARM64 (Graviton), ce qui nécessite que les images Docker soient compilées spécifiquement pour l'architecture linux/arm64. Ce que change cette combinaison est significatif pour les équipes qui déploient des agents en production : elle élimine plusieurs problèmes récurrents liés aux architectures vocales temps réel, notamment la gigue audio, les contraintes de montée en charge, et les coûts liés au sur-provisionnement. La facturation à l'usage actif — et non à la capacité réservée — réduit directement les coûts d'infrastructure inactive. Sur le plan technique, la latence reste le défi central : une conversation naturelle exige une réponse inférieure à une seconde de bout en bout. Pour y parvenir, le système mise sur le streaming bidirectionnel à deux niveaux — entre le client et l'agent d'une part, et entre l'agent et les modèles de langage d'autre part. Le choix du modèle est déterminant : AWS recommande Amazon Nova Sonic pour les pipelines speech-to-speech, ou Nova Lite dans une approche en cascade (STT → LLM → TTS), tous deux optimisés pour minimiser le Time-to-First-Token. La plateforme intègre également de l'observabilité native pour tracer le raisonnement de l'agent et ses appels d'outils. Ce premier volet d'une série de publications s'adresse aux développeurs déjà familiers des architectures vocales en cascade et speech-to-speech. Il fait suite à un article précédent d'AWS comparant Amazon Nova Sonic aux approches en cascade, et pose les bases techniques pour les déploiements Pipecat sur AgentCore Runtime.
This post is a collaboration between AWS and Pipecat . Deploying intelligent voice agents that maintain natural, human-like conversations requires streaming to users where they are, across web, mobile, and phone channels, even under heavy traffic and unreliable network conditions. Even small delays can break the conversational flow, causing users to perceive the agent as unresponsive or unreliable. For use cases such as customer support, virtual assistants and outbound campaigns, a natural flow is critical for user experience. In this series of posts, you will learn how streaming architectures help address these challenges using Pipecat voice agents on Amazon Bedrock AgentCore Runtime . In Part 1, you will learn how to deploy Pipecat voice agents on AgentCore Runtime using different network transport approaches including WebSockets, WebRTC and telephony integration, with practical deployment guidance and code samples. Benefits of AgentCore Runtime for voice agents Deploying real-time voice agents is challenging: you need low-latency streaming, strict isolation for security, and the ability to scale dynamically to unpredictable conversation volume. Without an appropriately designed architecture, you can experience audio jitter, scalability constraints, inflated costs due to over-provisioning, and increased complexity. For a deeper dive into voice agent architectures, including cascaded (STT → LLM → TTS) and speech-to-speech approaches refer to our previous post, Building real-time voice assistants with Amazon Nova Sonic compared to cascading architectures. Amazon Bedrock AgentCore Runtime addresses these challenges by providing a secure, serverless environment for scaling dynamic AI agents. Each conversation session runs in isolated microVMs for security. It auto-scales for traffic spikes, and handles continuous sessions for up to 8 hours, making it ideal for long, multi-turn voice interactions. It charges only for resources actively used, helping to minimize costs associated with idle infrastructure. Pipecat, an agentic framework for building real-time voice AI pipelines, runs on AgentCore Runtime with minimal setup. Package your Pipecat voice pipeline as a container and deploy it directly to AgentCore Runtime. The runtime supports bidirectional streaming for real-time audio, and built-in observability to trace agent reasoning and tool calls. AgentCore Runtime requires ARM64 (Graviton) containers, so make sure your Docker images are built for the linux/arm64 system. Streaming architectures for voice agents on AgentCore Runtime This post assumes your familiarity of common voice agent architectures: specifically the cascaded models approach, where you connect speech-to-text (STT) and text-to-speech (TTS) models in a pipeline, and the speech-to-speech model approach, like Amazon Nova Sonic . If you are new to these concepts, start with our earlier blog posts on the two foundational approaches: cascaded and speech-to-speech before continuing. When building voice agents, latency is a critical consideration, determining how natural and reliable a voice conversation feels. Conversations require near-instant responses, typically under one second end-to-end, to maintain a fluid, human-like rhythm. To achieve low latency, you need to consider bi-directional streaming on multiple paths, including: Client to Agent: Your voice agents will run on devices and applications, from web browsers and mobile apps to edge hardware, each with unique network conditions. Agent to Model: Your voice agents rely on bidirectional streaming to interact with speech models. Most speech models expose real-time WebSocket APIs, which your agent runtime or orchestration framework can consume for audio input and text or speech output. Model selection plays a key role in achieving natural responsiveness. Select models like Amazon Nova Sonic (or Amazon Nova Lite in a cascaded pipeline approach) that are optimized for latency and provides a fast Time-to-First-Token (TTFT). Telephony: For traditional inbound or outbound calls handled through contact centers or telephony systems, your voice agent must also integrate with a telephony provider. This is typically achieved through a handoff and/or Session Interconnect Protocol (SIP) transfer, where the live audio stream is transferred from the telephony system to your agent runtime for processing. In Part 1 of this series, we will focus on the Client to Agent connection and how to minimize the first-hop network latency from your edge device to your voice agent and explore additional considerations in relation to other components of voice agent architecture. To illustrate these concepts, we will explore four network transport approaches with considerations for: How users interface with your voice agents (web/mobile applications or phone calls) Performance consistency and resilience across variable network conditions Ease of implementation Approach Description Performance consistency Ease of
