Lab overview

Build and extend an AI agent

Tailwind Traders has a language model, but it has never read the company’s own documents — so it answers fluently and gets it wrong. In this lab you turn that model into a real assistant: ground it in your documents, give it tools it can call, and put your own app in front of it. Two core tasks get you a working, tool-using agent in about 35 minutes; four optional tasks take it all the way to a hosted agent.

You write

Python on your machine

A chat window is provided — you write the logic behind it.

client.pyTask 3

Loads the agent and routes each tool call

Your own toolsTasks 4–5

Python functions, then your own MCP server

You build

One agent in Microsoft Foundry

Four pieces — each task adds or upgrades exactly one.

InstructionsTask 1

Who it is, and how it should answer

KnowledgeTask 1

Your uploaded documents, indexed for file search

ToolsTasks 2, 4, 5

Remote MCP, your functions, your own server

HostingTask 6

Your code moves into a Foundry-run container

It reads

Real sources, not guesses

Where every grounded answer actually comes from.

Your documents

Store_Policy.txt, weekly_sales.csv

Learn Docs MCP serverTask 2

Live documentation, fetched at question time

Knowledge makes it accurate Tools make it useful A client makes it real Hosting makes it yours to run

Lab overview

Build and extend an AI agent

Ask a plain language model about your own company and it will answer confidently, and wrongly. This lab closes that gap: you ground an agent in real documents, then give it tools it can call so it can reach past them. Two core tasks, about 35 minutes — and the same question gets a very different answer.

Before · a model on its own

“What’s your return policy for a tent?”

Tents can be returned within 90 days, no receipt required, for a full refund.

✗ Invented — nothing here came from Tailwind

The model is not broken. It has simply never read the store policy, so it fills the gap with something plausible.

After · a grounded, tool-using agent

“What’s your return policy for a tent?”

Camping equipment may be returned within 30 days with proof of purchase, provided it is unused and in its original packaging.

✓ Answered from your own document

Grounded in Store_Policy.txt — and for anything outside your files, it calls the Learn Docs MCP server live.

What you add to get there: Knowledge Task 1 Tools Task 2 Your own app Tasks 3–4 Hosting Tasks 5–6

Lab overview

Build and extend an AI agent

One assistant for Tailwind Traders, built in four layers — and every task in this lab adds exactly one of them. You start with a model that guesses and finish with an agent that reads your documents, calls your tools, and runs inside Microsoft Foundry. Stop after any layer and you still have something that works.

Instructions + knowledge

Tell it who it is, upload the store policy, turn on file search — all in the portal, no code

AccurateTask 1 · ~15 min

Tools it can call

A remote MCP server for live docs, then your own Python functions

UsefulTasks 2, 4, 5

A client app

Your own chat window, talking to the agent you just built

RealTask 3

Hosting

Your code moves into a Foundry-run container with its own endpoint

Yours to runTask 6

Layers 1–2 are the core path — about 35 minutes to a working, tool-using agent.

Lab overview

Build and extend an AI agent

Turn a model that guesses into an assistant that answers from Tailwind Traders’ own documents, calls tools, and runs in Microsoft Foundry. Every task stands alone and shares one starter folder, so you choose how far to go — and you have something working after the first 35 minutes.

Core

Task 1

Accurate

Answers from your uploaded documents instead of guessing. Portal only — no code.

~15 min

Core

Task 2

Useful

Calls a remote MCP server for live documentation, and asks you before it acts.

~20 min · stop here and you have a working agent

Optional

Tasks 3–4

Real

Your own chat app drives it, and your own Python functions become tools it can call.

~45 min

Optional

Tasks 5–6

Yours to run

Host your own MCP server, then promote the whole thing to a hosted agent running in Foundry.

~65 min

Lab overview · application architecture

Build and extend an AI agent

You write Python locally; the agent, its knowledge and its tool definitions live in a Microsoft Foundry project. Each badge marks the task that builds that component — Task 1 is portal-only, everything else is code you run yourself, until Task 6 moves your code into Foundry too.

Your machine

tailwind_ui.pyprovided

Web chat window — you never edit it

client.pyTask 3

You write respond() and route each tool call

functions.pyTask 4

Trip planner and rental pricing, in Python

server.pyTask 5

Your own MCP server, tools over stdio

Microsoft Foundry project

Model deploymentsetup

The model your agent runs on

Agent · instructionsTask 1

Who it is, and how it should answer

Knowledge · File searchTask 1

Your uploaded documents, indexed

Tool definitionsTasks 2, 4, 5

Remote MCP, your functions, your MCP server

Agent ServiceTask 6

Your code in a hosted container, own endpoint

Data & remote sources

Your documentsTask 1

Store_Policy.txt, weekly_sales.csv

Learn Docs MCP serverTask 2

learn.microsoft.com/api/mcp · approval required

Task 1 Accurate — grounded in your files Task 2 Useful — reaches live sources Tasks 3–4 Real — your app, your Python Tasks 5–6 Yours to run — your server, hosted

Lab overview · what each task builds

Build and extend an AI agent

Six tasks, one Tailwind Traders assistant — each row is the component you add and what it buys you. Tasks share one starter folder and one .env, so you can run just one, or work straight through.

Task	Runs where	Component you build	Achievement
Task 1Core · ~15 min	Foundry portal	Agent + instructions File search index Store_Policy.txt	Accurate
Task 2Core · ~20 min	Foundry + local	Remote MCP tool Approval loop	Useful
Task 3Optional	Your machine	client.py Responses API · agent_reference	Real
Task 4Optional	Your machine	functions.py Function-calling loop	Extensible
Task 5Optional · capstone	Your machine	server.py · your MCP server stdio · JSON-RPC	Your own tools
Task 6Optional	Foundry	Agent Service container Own endpoint	Yours to run

Lab overview · one question, end to end

Build and extend an AI agent

This is a single question travelling through the assistant you build — every hop is a component, and every badge is the task that adds it. Build the first two hops and you already have a grounded, tool-using agent; the rest put your own code on both ends.

client.py

Your chat app sends the question and routes whatever comes back. Task 3

→ Responses API · agent_reference

Foundry agent

Instructions decide how it answers, and which tool it needs. Task 1

→ file search, or a tool call

File search

Looks the answer up in your uploaded documents, and cites them. Task 1

Store_Policy.txt · weekly_sales.csv

Tools

Not in your files? It calls the Learn Docs MCP server, or your own Python. Tasks 2, 4, 5

approval required before it acts

Hosted agent

Finally the same code runs inside Foundry, behind its own endpoint. Task 6

callers change nothing

Hops 2–3 are the core path — about 35 minutes Hops 1, 4, 5 are optional, in any order Scenario: Tailwind Traders

Lab overview · component view

Build and extend an AI agent

Three subsystems, wired by interfaces you either provide or require — each part tagged with the task that builds it. Task 1 needs no code at all; Task 6 relocates your own parts into the Foundry subsystem.

Lab overview · deployment view

Build and extend an AI agent

Where each artifact actually runs. Everything you write starts on your own machine and talks to a Foundry project over HTTPS — until Task 6 redeploys your code into Foundry, behind its own endpoint.

«device»

Your machine

tailwind_ui.py

provided chat window

given

client.py

calls the agent, routes tools

Task 3

functions.py

your Python tools

Task 4

server.py

your MCP server (stdio)

Task 5

«execution environment»

Microsoft Foundry project

Model deployment

the model the agent runs on

setup

Agent + knowledge

instructions, file search index

Task 1

Tool definitions

MCP, functions, approval

Tasks 2,4,5

Agent Service container

your code, moved here

Task 6

«external system»

Outside your tenant

Learn Docs MCP

called at question time

Task 2

Your documents

uploaded, then indexed

Task 1