building-ai-agent-on-cloudflare

To build stateful AI agents on Cloudflare Workers, use the Agents SDK with Durable Objects for persistent per-agent state and WebSocket communication. It supports tool integration and SQL storage for multi-step workflows and scheduled tasks.

Connect a client to a Cloudflare edge agent by targeting the WebSocket endpoint at wss://<your-subdomain>.workers.dev/agents/MyAgent/session-id. This enables real-time chat and tool-calling interactions directly through the Agents SDK.

Deploying edge AI agents requires Node.js 18+, Wrangler installed, a Cloudflare account, and the Agents SDK bindings. Optional AI bindings are needed for model calls, supporting edge deployments with per-agent routing.

Yes, you can use Durable Objects for persistent state in Cloudflare AI agents. The Agents SDK leverages Durable Objects and SQL storage to maintain per-agent state, enabling orchestrated multi-step workflows and scheduled tasks.

The best way to orchestrate multi-step workflows for edge AI agents is using the Cloudflare Agents SDK. It provides edge orchestration across components, tool integration for external services, and supports scheduled tasks with real-time WebSocket communication.

Scaffold a new Cloudflare agent project by running npm create cloudflare@latest -- my-agent --template=cloudflare/agents-starter. After running npm install and npm start, the local dev server launches for immediate edge deployment testing.