Summary
Giselle van Dongen from Restate delivers a production-focused talk on the infrastructure challenges of running AI agents reliably at scale, framed around Andrej Karpathy’s three-wave model of LLM evolution toward persistent, asynchronous agentic entities. The talk introduces Restate, an open-source durable execution framework inspired by Apache Flink and Meta’s event infrastructure, designed to handle long-running, stateful, distributed agent processes without requiring developers to hand-roll retry logic or recovery systems.
The core demo features a Slack-connected deep research agent. Van Dongen walks through Restate’s architecture — a server that acts as a proxy and lifeline between requests and agent services, journaling events to enable crash recovery at the exact point of failure rather than from scratch. Key primitives covered include durable promises (for human-approval workflows that may suspend for weeks), virtual objects (stateful actor-style sessions with isolated key-value state), and concurrency controls that prevent multiple executions from corrupting shared session state.
Practical patterns demonstrated include how to cancel or kill a stuck agent mid-execution, how to inject follow-up context into a running agent without waiting for it to finish, and how serverless deployments avoid burning execution time during long suspension windows. The talk is aimed at engineers moving from single-agent prototypes to production agentic platforms spanning multiple services and teams.
📺 Source: AI Engineer · Published September 14, 2026
🏷️ Format: Deep Dive
