Examples for AI workflows and Durable Agents

This repo contains a set of runnable examples of AI workflows and agents, using Durable Execution and Orchestration via Restate (Github)

The goal is to show how you can easily add production-grade resilience, state persistence, retries, suspend/resume, human-in-the-loop, and observability to agentic workflows. So you can ship agents that stay alive and consistent without sprinkling retry-code everywhere and without building heavyweight infra yourself.

The Restate approach works independent of specific SDKs but integrates easily with popular SDKs, like the Vercel AI SDK or the OpenAI Agent SDK. You can also use without and Agent SDK (roll your own loop) or for more traditional workflows.

Why Restate?

📄 For a gentle intro, read the blog post "Durable Agents - Fault Tolerance across Frameworks and without Handcuffs"

Use Case	What it solves
Durable Execution	Crash-safe LLM/tool calls & idempotent retries—agents resume at the last successful step.
Detailed Observability	Auto-captured trace of every step, retry, and message for easy debugging and auditing.
Human-in-the-loop & long waits	Suspend while waiting for user approval or slow jobs; pay for compute, not wall-clock time.
Stateful sessions / memory	Virtual Objects keep multi-turn conversations and other state isolated and consistent.
Multi-agent orchestration	Reliable RPC, queuing, and scheduling between agents running in separate processes.

Restate UI showing an agent execution with parallel tool calls

Full Example Catalog

Agent SDK Integrations

Integration	Example	Description	Code	Docs
Vercel AI SDK	Template	A minimal example of how to use Restate with the Vercel AI SDK		📖
	Tour of Agents	A step-by-step tutorial showing how to build resilient agents		📖
	Examples	More advanced examples that can be deployed as a Next.js app on Vercel		-
OpenAI Agents SDK	Template	A minimal example of how to use Restate with the OpenAI Agents SDK		📖
	Tour of Agents	A step-by-step tutorial showing how to build resilient agents		📖

Composable AI Patterns

Pattern	Description	Code	Docs
Chaining LLM calls	Build fault-tolerant processing pipelines where each step transforms the previous step's output		📖
Tool routing	Automatically route requests to tools based on LLM outputs		📖
Parallel tool execution	Execute multiple tools in parallel with durable results that persist across failures		📖
Multi-agent routing	Route requests to specialized agents based on LLM outputs		📖
Remote agent routing	Deploy/scale agents separately and route requests with resilient communication		📖
Parallel agent processing	Run multiple, specialized agents in parallel and aggregate their results		📖
Racing agents	Race multiple agents against each other and use the fastest response		📖
Human-in-the-loop pattern	Implement resilient human approval steps that suspend execution until feedback is received		📖
Chat sessions	Long-lived, stateful chat sessions that maintain conversation state across multiple requests		📖
Orchestrator-worker pattern	Break down complex tasks into specialized subtasks and execute them in parallel		-
Evaluator-optimizer pattern	Generate → Evaluate → Improve loop until quality criteria are met		-

Other Examples

Example	Description	Code
MCP	Using Restate for exposing tools and resilient orchestration of tool calls
A2A	Implement Google's Agent-to-Agent protocol with Restate as resilient, scalable task orchestrator

Restate currently supports 6 languages:

The examples can be translated to any of the supported languages. Join our Discord/Slack to get help with translating an examples to your language of choice.

Learn more

• Documentation
• Examples on workflows, microservice orchestration, async tasks, event processing
• Restate Cloud
• Discord / Slack

Acknowledgements

• The DIY patterns are largely based on Anthropic's agents cookbook.
• Some of the A2A examples in this repo are based on the examples included in the Google A2A repo.