Build, isolate and recover a grounded AI support assistant with Cloudflare Agents, Workers AI, Vectorize and AI Gateway. Connect retrieval, identity, durable state and approved actions into a working application.
The course contains three independent project challenges: launch a source-backed assistant, repair a customer-isolation leak and recover an approved action after a model-route failure.
What You Will Learn
- Launch a support assistant backed by retrievable sources
- Keep customer identity and knowledge retrieval inside the correct scope
- Repair a cross-customer knowledge leak
- Connect durable Agent state to explicit action approval
- Recover an approved action when the primary model route fails
- Observe application results and clean up disposable resources
Who This Course Is For
Learners with the Cloudflare AI and Agents foundations who want to test an integrated assistant independently.
Prerequisites: Complete Agents SDK, Workers AI, Vectorize and AI Gateway with their Workers and Durable Objects foundations. Direct-entry learners first complete Connect LabEx to Your Cloudflare Account.
Learning environment: Each challenge starts in a fresh LabEx VM with your own authorized learning account. Small synthetic datasets and uniquely named disposable resources keep the focus on identity, retrieval, approval and recovery.
Frequently Asked Questions
What does RAG add to this assistant?
Retrieval supplies source passages before generation so the assistant can support an answer with relevant evidence. The launch challenge connects that source-backed flow.
Is customer isolation only a prompt instruction?
No. The isolation challenge checks application identity and retrieval scope. Repair the boundary without mixing customers’ knowledge.
Can fallback skip action approval?
The recovery challenge preserves approved-action behavior when the primary model route fails. Recovery must keep the identity and approval boundaries in place.
Do I reuse the previous challenge’s Agent?
No. Each challenge starts independently. Reuse knowledge and your account, verify actual application results, delete only the resources you created and log out after cleanup.
