4.3. Reflection Checkpoint

Key Takeaways

• The two-client pattern is everything: AIProjectClient (setup, via endpoint + DefaultAzureCredential) then get_openai_client() (does the work). Recognize it and you can read most Foundry code.
• Deploy before you call. Models come from the catalog, get a deployment name, and are testable in the playground before any code. Code references the deployment name.
• System prompt vs. user prompt: durable rules go in the system prompt; specific requests go in the user prompt. Specific, contextual prompts beat vague ones.
• An agent = model + instructions + tools, and its hallmark is autonomous action toward a goal. Use a plain model call for generation; use an agent when the task requires taking actions.
• Agent code adds two things to the chat pattern: a conversation (multi-turn memory) and an agent_reference (route through the agent).

Connecting Forward

You can now deploy a model, prompt it well, and build a single agent. Phase 5 widens the toolkit to other modalities: building a text-analysis application, interpreting images with a multimodal model, handling speech recognition and synthesis, responding to spoken prompts, and generating images. The same two-client pattern underlies all of it — you're mostly changing what you send in and what you do with what comes back.

Self-Check Questions

• Walk through, from memory, the three moves of a minimal Foundry chat client. Then name the two additions that make it an agent client.
• A teammate's agent ignores its "only discuss geography" rule on the second message. Given what you know about instructions and conversations, name two things you'd check.