Build stateful AI agents that hibernate between calls

What it does

cloudflare/agents is a TypeScript framework for deploying persistent, stateful AI workloads on Cloudflare Workers. Each agent lives inside its own Durable Object with SQLite storage and a lifecycle, letting you run millions of isolated instances that wake on demand and sleep while idle.

The interesting bit

Instead of treating AI agents as ephemeral serverless functions, the SDK gives each one a long-running identity at the edge. It bundles a broad stack into the same runtime: real-time sync, scheduling, MCP plumbing, voice pipelines, sandboxed code execution, and even pay-per-call APIs via the x402 protocol.

Key highlights

• Core Agent class handles state persistence, @callable() RPC, WebSockets, and scheduling out of the box.
• Specialized packages add voice pipelines (@cloudflare/voice), LLM-generated code execution (@cloudflare/codemode), and sandboxed Workers (@cloudflare/shell).
• Supports the Model Context Protocol as both client and server, plus WebMCP to tunnel browser-side tools over WebSocket.
• Includes React hooks (useAgent, useAgentChat, useVoiceAgent) and a vanilla AgentClient for frontend sync.
• Over 30 self-contained examples, from chat and workflows to tic-tac-toe and email agents.

Caveats

• The experimental/ directory contains work-in-progress features with explicit “no stability guarantees.”
• Several older AI-chat entry points are deprecated and still exist only as re-exports, so check import paths against the current package list.

Verdict

Worth exploring if you already run on Cloudflare and need persistent, stateful agent instances rather than stateless inference wrappers. Skip it if you are not prepared to adopt the Durable Object and Workers ecosystem lock-in.