agents-sdk

Install the agents package, extend the Agent class with typed state and an initialState property, then configure a Durable Object binding and a new_sqlite_classes migration in wrangler.jsonc. Route requests with routeAgentRequest and expose methods to clients using the @callable decorator.

Use this.schedule() for one-time delayed, date-based, or cron tasks, and this.scheduleEvery() for fixed intervals. The handler method receives the payload and schedule object; cron schedules auto-reschedule while one-time schedules are deleted after running.

AIChatAgent requires you to write the streamText loop and wire tool execution yourself inside onChatMessage, giving full control. Think handles the LLM loop, tool execution, and persistence automatically via getModel() and getSystemPrompt(), but requires the experimental compatibility flag.

Yes, agents include a multi-server MCP client. Call this.addMcpServer() with the server URL, then use this.mcp.getAITools() to merge MCP tools into your chat agent. You can also build MCP servers using the McpAgent class with streamable HTTP transport.

Enabling experimentalDecorators in tsconfig.json conflicts with the decorators used by the Agents SDK and breaks @callable methods. Extend the agents/tsconfig configuration instead, which sets the correct compiler options for the SDK.

Use the built-in queue() for background tasks under 30 seconds with sequential FIFO processing. Use AgentWorkflow for long-running multi-step tasks, automatic retries, and human-in-the-loop approval flows that must survive failures.