Building Resilient AI Agents (www.restate.dev)

Restate announced Durable Execution and Restate Cloud — a system that makes serverless AI agents stateful, resilient, and production-ready without forcing you to run your own worker fleet. Instead of keeping Lambda/Vercel/Cloudflare functions running or building complex queues and orchestration logic, Restate records each meaningful step (LLM responses and tool executions) in a durable journal. If a function crashes or needs to wait for a human approval, the workflow is suspended via a durable promise, the function can scale to zero, and Restate later re-invokes the function to replay recorded results and resume exactly where it left off. Restate’s server sits in front of serverless functions (pushing requests to them rather than requiring worker pulls), and an SDK exposes run-blocks, durableCalls middleware, and promise primitives to mark durable steps. This matters because agent workloads are bursty, long-running, and non-deterministic: LLM outputs must be recorded for correct replay, tool side effects need persisted results, and in-flight executions must survive version upgrades. Restate addresses these with a distributed, event-driven engine that claims tens-of-thousands concurrency, live execution timelines for observability, and immutable deployment URLs so suspended executions resume on the original code. The managed Restate Cloud lets teams get these guarantees quickly, enabling cost-efficient human-in-the-loop flows, simpler integrations with AI SDKs (Vercel AI, OpenAI Agents, custom stacks), and safer evolution of agents in production.