Every step you take, every call you make: the reliable agent stack — Giselle van Dongen, Restate

A research agent asks a human to approve its plan, then waits. The wait might last a month, spanning restarts and redeploys, and while it waits the function consumes no serverless execution time at all. Giselle van Dongen builds Restate, and that suspended promise is the clearest illustration of what durable execution buys you. Her talk skips the agent loop itself and goes after the infrastructure underneath it, the part she argues teams keep rebuilding badly: retry logic, recovery logic, session isolation, and stopping something already in flight. Restate runs as a server in front of your agent service, proxying requests and holding an open connection that acts as a lifeline. As the agent works it emits events to a journal, and that journal replays the process back to its exact failure point rather than starting over. The design borrows from Apache Flink and Meta's event infrastructure, and it pushes invocations rather than polling, which is where the low latency comes from. The demo is a deep research agent living in Slack. A planner proposes subtopics, parallel sub agents go to work, and an injected tool failure shows a web search retrying and completing instead of sinking the run. Then it gets more interesting. Because a session is modeled as a virtual object with its own key and isolated state, van Dongen can talk to a run already in progress. She tells it midflight to focus on frontier models and a classifier decides that is relevant, so it signals the live loop. She then tells it to research something else entirely, and the cancellation rewinds down the call chain, killing sub agents before the controller. Speaker info: - https://x.com/vdgiselle - https://www.linkedin.com/in/giselle-van-dongen/ Timestamps: 0:00 - Three waves, and where agents are heading 1:53 - The infrastructure layer nobody wants to write 2:45 - Four ingredients of a durable foundation 4:29 - How the journal recovers a failed run 5:20 - Demo: a deep research agent in Slack 7:51 - Making a step durable in code 8:42 - Suspending a month for human approval 9:32 - Virtual objects as stateful actors 12:10 - Signaling a running agent loop 13:51 - Cancelling and rewinding the stack 17:19 - Inside the distributed log