Building the Network for AI Agents — Yan Avlasov, Google
Sep 14, 2026 · 54 min · 13 segments
AI agents bring new demands to the infrastructure connecting them to models, tools, and data. Each agent can have its own permissions and state, and spend much of its time asleep before waking up to…
AI agents bring new demands to the infrastructure connecting them to models, tools, and data. Each agent can have its own permissions and state, and spend much of its time asleep before waking up to make another request. Supporting that behavior at scale changes how the network routes traffic, loads policies, and manages resources. In this episode of Alexa’s Input (AI), I sit down with Yan Avlasov, Staff Software Engineer at Google and a senior maintainer of Envoy, to talk about building the networking stack for agents. We get into why Google is building on Envoy, what changes between serving inference and supporting agents, and why enforcing policies for individual agents requires a more dynamic control plane. Yan also shares what he’s seeing from AI-powered security scanning, including how models combine subtle bugs into serious vulnerabilities and why fixing them can mean changing behavior that production systems have relied on for years. From the episode: \- Model selection, cost controls, and fallback across providers \- Loading policies dynamically as agent traffic changes \- Managing personalized state and agents that suspend and resume \- Where Model Context Protocol (MCP) fits alongside the protocols agents already use \- Semantic routing and choosing models based on what a request needs \- Security findings involving policy enforcement, path normalization, and JSON parsing \- The engineering work and cost of making security scanning continuous Chapters 00:00 Introduction 01:12 Welcome, Yan 02:30 How AI is changing Google 07:56 Yan’s networking background 10:00 New networking layers for agents 13:11 AI gateways, cost controls, and fallback 16:38 Why Google builds on Envoy 22:21 Per-agent policies and a dynamic control plane 26:56 Inference versus stateful agents 31:58 Project Substrate: running agents at scale 33:31 MCP and the other protocols agents use 37:32 Semantic routing and model costs 40:40 AI security scanning and the work of fixing bugs 43:55 Policy timing, path normalization, and JSON 47:11 Making security scanning continuous 50:40 What’s next for agentic networking 53:32 Closing
AI agents bring new demands to the infrastructure connecting them to models, tools, and data. Each agent can have its own permissions and state, and spend much of its time asleep before waking up to…
