A conventional Worker can answer a request, but many applications also need a stable identity that remembers state, coordinates connected clients and continues work after one request ends. The Cloudflare Agents SDK provides that application model on top of Workers and SQLite-backed Durable Objects.
This course builds a small synthetic support system one boundary at a time. You begin without an AI model so named identity, durable state and synchronized clients remain visible. Later labs add persistent conversation, scheduled work, model-selected tools, explicit approval, a read-only MCP surface and a practical routing diagnosis. Each new concept appears beside the application need that motivates it, so previous Agents SDK experience is not required.
The course contains 8 guided labs and 1 independent approval-repair challenge. Guided Labs introduce each capability before the independent challenge.
What You Will Learn
- Create a Named Support Agent: route stable names, persist support state and prove isolation between Agent instances.
- Synchronize a Support Dashboard: connect two browser clients and observe server-authoritative state updates in real time.
- Stream a Persistent Conversation: stream a bounded answer and restore durable conversation history after reconnect.
- Schedule a Support Follow-Up: register, inspect, complete and cancel persistent Agent tasks.
- Add Validated Support Tools: connect model-selected tools to synthetic records with schemas and server-side scope checks.
- Require Approval for Record Changes: pause a side effect for a human decision and make repeated delivery safe.
- Publish a Read-Only MCP Tool: expose a strictly validated synthetic lookup through a remote MCP transport.
- Diagnose a Misrouted Agent Session: repair the selected Agent name without losing history or weakening isolation.
- Unapproved Support Update: independently repair approval-before-effect and exactly-once behavior in a supplied Agent.
Who This Course Is For
Developers familiar with Workers and Durable Objects who want to build stateful assistants and coordinated clients.
Prerequisites: Complete Cloudflare Durable Objects for Beginners and Connect LabEx to Your Cloudflare Account. You should recognize a small TypeScript Worker and the idea of one Durable Object per stable name; this course introduces the Agents SDK itself from the beginning.
Learning environment: Use your own dedicated Cloudflare learning account. Every unit starts in a fresh VM, authorizes Wrangler separately, creates uniquely named disposable resources and finishes by deleting them and logging out. The exercises stay within bounded Workers Free resources; model-enabled labs use only a small number of Cloudflare-hosted inference calls.
Frequently Asked Questions
Do I need previous Agents SDK or AI experience?
No previous Agents SDK experience is required. The first two labs use no AI model so you can understand named identity, durable state and synchronized clients first. Model-enabled labs require the Workers AI foundations.
Can the model perform a record change directly?
A proposed tool call must pass schema and server-side scope checks. The approval lab pauses a side effect for a human decision and makes repeated delivery safe; the challenge checks approval before the effect.
What happens when a client reconnects?
The labs distinguish the selected Agent identity, durable conversation and server-authoritative state from one browser connection. You also repair a misrouted session without discarding valid history.
How are resources cleaned up?
Delete only the lab’s disposable Workers and object namespaces while authorized, confirm removal and log out. Each VM authorizes your own account independently.
