Build Autonomous
AI Agents at Scale

Extend the base Agent class to build autonomous AI agents with built-in lifecycle hooks, HTTP handling, and WebSocket support.

Route requests to named Agent instances using getAgentByName. Agents are addressable, persistent micro-servers that run your code.

Wire up Agents via wrangler.toml/jsonc with Durable Object bindings, SQLite migrations, and environment variables.

Persistent bidirectional connections with onConnect, onMessage, and onClose hooks. Stream live updates back to clients in real-time.

Handle standard HTTP requests via onRequest and stream incremental LLM responses or progress events to clients using SSE.

Automatically sync agent state to connected clients using setState and the useAgent React hook. Zero-config reactive state.

Call any LLM — OpenAI, Anthropic, Workers AI — using the AI SDK or native fetch. Stream tokens directly to WebSocket or SSE clients.

Embed documents into Cloudflare Vectorize, run similarity search, and inject relevant context into LLM prompts for accurate, grounded answers.

Every Agent has its own embedded SQLite database accessible via this.sql. Run queries, store structured data, and build memory into your agents.

Use this.setState and this.getState for fast key-value state storage. Changes auto-broadcast to all connected WebSocket clients.

Schedule one-time or recurring tasks with this.schedule using cron syntax or delay-based timing. Agents run exactly when needed.

Run stateful multi-step workflows that guarantee execution with automatic retries. Workflows survive failures and can run for hours or days.

Integrate headless browser services to let agents fetch, scrape, and interact with web pages as part of autonomous research or data-gathering tasks.

Sequential LLM calls where each step refines the previous output — with quality gates to trigger rework until standards are met.

Classify inputs and dynamically route to specialized downstream agents or models — simple queries to mini models, complex ones to powerful models.

Run multiple LLM calls simultaneously with Promise.all for independent sub-tasks. Aggregate specialist results for higher-quality output.

A planning LLM breaks work into sub-tasks and delegates to specialized worker LLMs. Results are synthesized into a final coherent output.

A generator LLM produces output while an evaluator LLM scores quality and provides feedback — iterating until the bar is cleared.

Build and deploy remote MCP servers on Cloudflare. Expose tools, resources, and prompts to any MCP-compatible AI client like Claude or Cursor.

Pause agent execution to request human approval or input before proceeding with sensitive or irreversible actions. Await confirmation via WebSocket or HTTP.

Enable agents to autonomously pay for API access using the x402 standard over HTTP 402. No accounts or credential management needed.