Durable Execution and AI Agent Infrastructure
Restate
Restate builds a durable execution runtime for AI agents, workflows, and backend services. Developers keep business logic in ordinary application code while Restate records progress and results, allowing execution to recover from failures without repeating completed durable steps. Its building blocks combine functions, reliable service calls, keyed state, queues, and persistent timers. For agent developers, these support durable model and tool calls, isolated sessions, coordination between agents, and waiting for human approvals without holding compute.
Co-founded by Stephan Ewen, Till Rohrmann, Igal Shilman, and Ahmed Farghal, Restate grew from experience with Apache Flink and transactional applications that did not fit analytical stream processing. Its log-centered architecture brings execution journals, communication, and state management into one system. A purpose-built replicated log persists new events; processors act on them and maintain state indexes, with snapshots stored in object storage. The runtime ships as a single binary, rather than requiring an external database or message broker.
Teams can self-host Restate, use fully managed Restate Cloud, or choose Restate BYOC, which runs managed infrastructure in their own cloud account and VPC. BYOC uses reserved-capacity pricing. In its 2026 announcement, the company reported production deployments at multiple customers, including setups exceeding 100,000 durable actions per second. Replit uses Restate for durable orchestration across its coding-agent platform.
Explore the recordings
The supplied Restate archive contains one recording. Use it as a guide to durable agent infrastructure and coordination; the capabilities and deployment options described below are claims made in the recording, not independently verified facts about Restate today.
Start with failure recovery and long-running work
In Every step you take, every call you make: the reliable agent stack, Giselle van Dongen explains infrastructure around agents: retries, recovery, and work that must survive interruptions. Follow the Slack deep-research demo for a concrete path through planning, parallel research, an injected web-search failure, and human approval. The recording describes journaled execution that resumes at a failed step and a durable promise that can wait through restarts or redeployments. This is an infrastructure talk; it does not cover agent-loop design or evaluations.
Then explore sessions, communication, and cancellation
Return to the same recording for coordination between running agents and users. The talk models sessions as virtual objects with isolated state and serialized execution, then demonstrates supplying context to an active research loop and canceling it when the user changes topics. The speaker describes cancellation propagating to sub-agents, model-call routing through an LLM gateway, and the underlying log, timers, and push-based invocation model. Treat its orchestration comparisons and self-hosted, BYOC, and managed deployment descriptions as recorded statements, rather than a current product checklist.
1 talk
Newest first
1 speaker at AIE
Affiliations reflect their AIE appearances, not necessarily current employment.