Restate

Restate provides durable execution — any async function can be made retryable and checkpointed transparently, without manual state management. For agents, it enables durable multi-step workflows where each step is automatically replayed from the last checkpoint on failure. Agent integration pattern: register service handlers with a Restate server; invoke them via HTTP ingress. Each invocation gets an idempotency key; Restate deduplicates retries automatically. Good fit for agents that need exactly-once processing semantics for critical workflows (payments, data pipelines, long-running LLM chains). Smaller community than Temporal; project is newer (2023). Confidence is docs-derived.

Restate Cloud uses an API key for service deployment registration. Ingress invocations support an x-restate-auth header for request authentication against registered handlers. Self-hosted instances require custom auth configuration via ingress middleware. Handler identity verification uses HMAC signing for Restate-to-service callbacks. Restate Cloud tenant isolation enforced at the server level.

docs.restate.dev provides SDK guides for all supported languages, concept explanations (durable execution, journals, timers), deployment guides for self-hosted and Restate Cloud, and API references. The concepts documentation is notably clear for understanding the execution model. Getting started requires running a Restate server (Docker image available) or signing up for Restate Cloud. Community support via Discord.

HTTP ingress at <restate-server>/invoke/{ServiceName}/{HandlerName}. Handlers are defined using the Restate SDK (TypeScript, Java, Kotlin, Go, Python) and registered with the Restate server. Invocations can be synchronous (wait for result) or one-way (fire and forget). Delayed calls supported via Restate's timer service. REST admin API for deployment registration, introspection, and cancellation. Response is the handler's return value serialized as JSON.

Failed handler invocations are automatically retried with configurable backoff until success or explicit terminal failure. Error propagation: handlers throw terminal errors to stop retries; other exceptions trigger retry. Restate's journal-based replay ensures no duplicate side effects on retry. Restate Cloud status at status.restate.dev. Self-hosted reliability depends on deployment configuration (single-node vs. distributed). Observability via OpenTelemetry traces.