2.2.1. Agent Components: Model, Instructions, Tools, Knowledge
💡 First Principle: These four components map one-to-one to the four things a completion lacks. The model is the reasoning engine (a deployment), instructions are the durable system prompt, tools are the actions it can request, and knowledge is the data it grounds on. Configure the right component and most "how do I make the agent do X" questions answer themselves.
Creating an agent binds these together. Instructions define the role and constraints (like a system message that lives with the agent, not just one request). Tools are registered capabilities — the model decides when to use them. Knowledge sources (an Azure AI Search index, uploaded files, Bing grounding) supply grounding data the agent retrieves from.
agent = project.agents.create_agent(
model="my-gpt4o-deployment",
name="support-agent",
instructions="You help customers using only the connected product manuals. Cite sources.",
tools=[code_interpreter_tool, ticket_lookup_tool],
)
| Component | Maps to | Example | Configured where |
|---|---|---|---|
| Model | Reasoning engine | gpt-4o deployment | Deployment (project) |
| Instructions | Durable behavior | "Answer only from manuals" | Agent |
| Tools | Actions it can request | ticket lookup, code interpreter | Agent |
| Knowledge | Grounding data | AI Search index, files | Agent (via connection) |
⚠️ Exam Trap: Instructions and knowledge are different levers. "The agent should follow our tone" is instructions; "the agent should answer from our docs" is knowledge (grounding). Conflating them — e.g., pasting documents into instructions — wastes context and doesn't scale, and is a wrong-lever answer.
Reflection Question: A requirement says the agent must "always refuse to discuss competitors and always answer from the current price list." Which clause is an instructions concern and which is a knowledge concern?
