Production agents need more than model calls and memory. Long-running systems also need durable progress, retry and recovery behavior, isolated state for concurrent sessions, communication between agents and tools, and controls that can cancel or redirect work already in progress.
The demonstration uses a Slack-connected research workflow to show how an event journal can restore progress after tool failures instead of restarting the entire run. Durable steps can also suspend without consuming compute while waiting for human approval, then resume from the recorded point when a response arrives.
Persistent sessions are modeled as stateful virtual objects with isolated histories and serialized execution. Signals can inject new context or cancel a running call chain, while shared gateways can add policy checks and concurrency limits around model access. Underneath, an event-driven distributed log persists journal events, state, timers, and inter-service requests.
Watch on YouTube
