AI Agents That Run a Business: Microsoft Foundry + Claude

Build four AI agents that run a small firm's back office - inbox, invoices, collections, weekly report - on Microsoft Foundry with MCP and Claude, then measure cost, quality and safety.

You sign in once to start.

11: modules
45: video lessons
10: quizzes
5.6 h: of video

0:00 / 0:00

24 lectures
20 lab videos
1 demos
10 quizzes

Each bar is one lesson. Gaps separate the modules.

What you will be able to do

Build an AI agent in the Microsoft Foundry portal and again in Python with the Microsoft Agent Framework
Give agents real tools through an MCP server you write yourself
Turn supplier invoices into structured data with a human approval step
Run the Claude Agent SDK inside a Foundry project for multi-step reasoning
Schedule an agent with a routine and post its report to Teams through a Logic App
Evaluate agents with Foundry evaluations and traces
Show cost per task and compare models before paying for one
Secure, test and monitor agents before handing them to a business

Curriculum

55 items (lessons and quizzes) across 11 modules. Pick any lesson to open it. You will sign in first if you have not already.

Course Introduction

5 lessons

Module 1: Harbor Accounting and the plan

6 lessons

Module 2: First agent in thirty minutes

5 lessons

Module 3: The same agent as code

5 lessons

Module 4: Tools via MCP

5 lessons

Module 5: The Invoice agent

5 lessons

Module 6: The Collections agent with Claude

5 lessons

Module 7: The Monday Report agent

5 lessons

Module 8: Evaluation and cost

5 lessons

Module 9: Ship it safely

8 lessons

Module 10: Putting it together

1 lesson

About this course

Build four AI agents that run a small firm's back office on Microsoft Foundry, give them real tools through an MCP server you write, bring in Claude for the reasoning job, and measure every one of them in seconds, cents and what a person still checks.

Harbor Accounting is a six-person bookkeeping firm with eight clients and one shared inbox. Tom spends ninety minutes a day sorting about sixty emails. Sana re-types twenty-five supplier invoices a week, twelve minutes each. Lena loses her Friday afternoons chasing eighteen overdue invoices. Priya, the owner, builds the Monday numbers by hand every Sunday evening. You join as the person who makes the computers help, and over eleven modules you build the agent team that gives those hours back: Triage, Invoice, Collections and the Monday Report. Agents draft, humans send, and an agent only has the identity and the tools its job needs.

What you build

Triage, three ways
first as a prompt agent in the Foundry portal, then as Python with the Microsoft Agent Framework deployed as a hosted agent, run on forty real emails and compared with the portal version.
The Harbor MCP server
one tool server for invoices and clients, run in a container, attached to Foundry agents through a toolbox that asks for approval before anything is written, and used again from Claude and Copilot.
Invoice and Collections
an Invoice agent that turns supplier PDFs into JSON, with a threshold in code that sends anything over five hundred dollars to a person, and a Collections agent on the Claude Agent SDK that drafts every chase email into an approval queue and never sends one.
The Monday Report, then measured and secured
a hosted report agent run by a Foundry routine that posts to Teams through a Logic App; Foundry evaluations, traces, a cost meter and model routing; then agent identities, Key Vault, a prompt shield and a poisoned-invoice test that your own agents catch twice, with a runbook, an approvals table and a switch-off section for the client.

How every module works

Pain
thirty seconds on what hurts at Harbor, in numbers.
Build
step by step on screen, with the exact portal labels and repo files.
Run
on the same generated Harbor dataset every time, so your numbers match the video.
Measure
one line at the end, such as forty emails, about eighty seconds, one cent, three for a human, in the course's own tests.

What you get

5.6

hours of video

lectures

lab videos

portal demos

quizzes

Forty-five captioned video lessons in eleven modules, about five and a half hours, with real Foundry, Azure and VS Code screens and short animations
Nine lab guides with checkpoints and expected output, plus the full harbor-agents repo and the dataset generator
Ten short knowledge checks
identity page, approvals table, and a runbook with a switch-off section

You need basic Python and an Azure subscription with access to Microsoft Foundry. The Teams, Outlook and Claude steps are optional and shown in full. Claude in the Foundry catalogue needs pay-as-you-go billing, so the course calls Claude with an Anthropic key. Preview features, such as routines, trace evaluation and the Monitor tab alerts, are named as preview whenever they appear.