Pass 901

Owen Park

Episode eight. Agents. The skills list says: create and test a single-agent solution in the Foundry portal, and create a lightweight client application for an agent. Two surfaces, same idea.

Mira Cole

Definition you can reuse. An agent is a model paired with instructions and tools so it can reason over a request and take action. A chat completion answers. An agent can look things up and do steps.

Owen Park

Foundry talks about a spectrum. Declarative prompt agents live in the portal: you fill in name, model, instructions, tools, knowledge. Hosted agents are your own code in a container, more control, more work. For A.I. nine-oh-one, the portal single-agent path is the one to love.

Mira Cole

Instructions are the agent's system prompt wearing a badge. Who it is, what it may do, what it must refuse. Tools are functions or connected skills: get order status, search docs, send email. Knowledge is grounded content, files or indexes, so it is not inventing policy.

Owen Park

Function calling, said slowly. The model does not execute your Python. It emits a structured request: call get_order with order_id one two three. Your runtime, or Foundry's tool host, runs that function, returns the JSON, and the model then writes the user-facing sentence. If a question says the model itself queried S.Q.L., that answer is wrong.

Mira Cole

Testing in the portal means sample inputs. 'Where is order four two one?' You watch traces: did it call the tool, did it stay inside instructions, did it cite knowledge. Evaluation and tracing are how responsible A.I. shows up in the workshop.

Owen Park

Lightweight client for an agent is the episode-seven program with a different target. You still make an A.I. Project Client. Instead of chatting with a raw deployment name, you create or reference an agent version, then send a conversation to that agent. Packages still look like azure-ai-projects and azure-identity.

Mira Cole

Grounding and retrieval, one more time, because agents make it obvious. If the agent must answer from this week's price list, attach knowledge or a tool that fetches the list. Do not raise temperature and hope.

Owen Park

Single-agent is the exam scope. Multi-agent orchestras exist in longer learning paths. If a fundamentals question offers 'design a swarm of twenty agents with custom networking,' that is a distractor wearing a fancy coat.

Mira Cole

Question one. What does an agent have that a single chat completion usually does not?

Owen Park

Tools, and the ability to take multiple steps toward a goal. Instructions are similar to a system prompt. Tools are the difference that matters.

Mira Cole

Question two. In function calling, who executes the get_inventory function?

Owen Park

The application or the platform's tool host. The model only requests the call and later reads the result.

Mira Cole

Question three. You need a no-code prototype that answers from uploaded P.D.F.s and can look up tickets. Where do you start?

Owen Park

Create a single-agent solution in the Foundry portal. Attach knowledge and a ticket tool. Test with sample questions. Code can wait.

Mira Cole

Question four. A lightweight client should talk to an agent that already exists in the project. What is the high-level flow?

Owen Park

Authenticate, construct the project client, reference the agent by name or version, send the user message through the agent A.P.I., print the response. Do not re-deploy the whole catalog.

Mira Cole

Question five. An agent keeps inventing refund policies. What should you add first?

Owen Park

Ground it. Attach the real policy as knowledge, tighten instructions to answer only from that knowledge, and evaluate groundedness. A hotter temperature is how you get more fiction.

Mira Cole

Episode nine is the rest of the Foundry app catalog: text analysis apps, spoken prompts into multimodal models, Azure Speech in Foundry Tools, vision in prompts, and generating new images.

Owen Park

Bring ears and eyes. It is a multimodal episode.

Agents in Microsoft Foundry

On the mics