Nine production patterns, each as a clickable sequence diagram. Pick a pattern, then click any step to see what happens, what it protects against, and what happens when it fails. Purple steps are where a human approves, confirms, takes over or reviews. Switch the view to AWS, Google Cloud or Azure to see the services behind each step.
Same building blocks every time: input checks, context, a model, tools, output checks, feedback. What changes is who controls the flow and where the human sits.
| Pattern | Who controls the flow | Main risk | Key control | Where the human sits | Seen in |
|---|
The bill-pay example uses Hybrid + rerank for policy documents, plus a memory lookup. The live balance is not RAG: it comes from a tool call in steps 11–12. Pick a type to see how steps 5–6 change.
Every pattern keeps a person on the steps where a mistake is expensive or irreversible. Purple arrows in the diagrams mark these steps.
The agent pauses and shows the exact action. Nothing runs until the person confirms. Bill pay, purchases, risky shell commands.
The person sets scope up front, or takes control mid-task: CAPTCHA and logins in a browser, a human agent on a hard call.
Automation handles the confident cases; people see only what is uncertain or breaks a rule. Invoices, low-confidence form fields.
A person reviews before anything ships: PR review and merge, advisor approval, legal review of ads.
The building blocks match across clouds. Use the view switch above the diagram to see each cloud's service on every lane and step. Google renamed Vertex AI to Gemini Enterprise Agent Platform in 2026, and Azure AI Foundry is now Microsoft Foundry; the page uses the current names.
| Building block | AWS | Google Cloud | Azure |
|---|
Every pattern above runs on the same AWS building blocks. Amazon Bedrock is where you use models and managed agent parts. Amazon SageMaker AI is where you build, tune and host your own models. Switch the diagram view to AWS to see the service behind each lane and step.
Foundation models (Claude, Llama, Mistral, Amazon Nova and more) behind one API, plus managed building blocks.
Train, fine-tune, evaluate, host and monitor models you control: open-weight, fine-tuned or fully custom.
Rule of thumb: start on Bedrock. Move one step to SageMaker AI when you need custom weights, a model Bedrock doesn't host, specific hardware or latency, or full control of serving. Small custom models (injection classifier, reranker, embeddings) are the usual first SageMaker endpoints.
| Step in the flow | Amazon Bedrock | Amazon SageMaker AI | Around them |
|---|
“The model proposes. The orchestrator disposes. Every tool call passes a policy check, runs with the user's scoped credentials, needs human approval for irreversible actions, and gets a grounding check on the way out.”