1.5. Reflection Checkpoint

Key Takeaways

• Treat Foundry as a platform with a containment hierarchy (resource → project → deployment/agent/connection), not a renamed OpenAI resource — match each configuration concern to its layer.
• An agent = model + instructions + tools + knowledge + thread; reach for it when a requirement needs actions, multi-step work, or persistent state, and reach for a plain completion when it doesn't.
• Grounding supplies facts, fine-tuning supplies behavior, prompting supplies steering — diagnose what's missing before choosing a lever; "fine-tune on our docs to add knowledge" is a trap.
• Run the four decision filters (boundary → grounding → agent workflow → modality) on every scenario to discard true-but-irrelevant distractors.

Connecting Forward

Phase 2 dives into the heaviest domain — generative AI and agentic solutions — where these first principles become concrete: you'll connect to a Foundry project with the SDK, shape outputs with prompting and structured/JSON output, register tools and build agents in the Agent Service, orchestrate multiple agents, add multimodal generation, and evaluate output quality. Every one of those topics is an application of the agent definition and the decision filters you just built.

Self-Check Questions

• Without naming a single service, explain why AI-103 is "an assembly exam" and AI-102 was "a selection exam." What does that change about how you read a question?
• A colleague proposes solving "the model doesn't know our return policy" by switching to a larger model. Using first principles, explain in two sentences why that won't work and what will.