Microsoft Foundry & Python: Enterprise AI Agents on Azure

Explore AI agents and AI-driven workflows in the enterprise, learn how agents differ from automation, and see how they reason, plan, and act using Microsoft Foundry in real-world use cases.

Explore how AI agents combine brain (LLM), memory, and tools to act, using Microsoft Foundry and Python. Learn prompts, governance, and enterprise readiness to build scalable, observable agents.

Explore Microsoft Foundry and the agent framework architecture that unifies models, data, and tools in Azure to build, deploy, and operate enterprise AI agents.

Set up python 3.10 and Azure CLI prerequisites, create an Azure subscription, then deploy a model in Microsoft Foundry to wrap it as a reusable agent with tools and memory.

Install the Azure CLI with winget on Windows 11 for a one-command setup, verify with az --version, and log in to connect subscriptions to Azure and Foundry.

Set up Visual Studio Code for building AI agents and workflows with Microsoft Foundry and the AI Toolkit, install extensions, sign in, and connect Python apps to Azure-hosted agents.

Set up a clean python environment for Microsoft Foundry and Azure AI Foundry, using a virtual environment, activate automatically, install required packages, and authenticate with the Azure CLI.

An agent resides in Azure; a Python client calls it through the Foundry project endpoint, using the AI Project Client and Default Azure Credential to authenticate.

Compare the Azure AI Projects SDK and the Microsoft Agent Framework SDK, highlighting simple sync code versus async-first lifecycles with tools, streaming, memory, and advanced workflows.

Prototype AI agents fast using the AI Toolkit VS Code extension, exploring the agent builder, playground, evaluation tab, and variables in a sandbox, Foundry handles hosting, orchestration, monitoring, and scaling.

Prototype an artificial intelligence agent that turns ingredients into a recipe, test locally in a sandbox, refine prompts with generate and playground, then deploy to Foundry and Azure.

Explore how Microsoft Foundry unifies agents, models, and tools on Azure to design, deploy, and manage production-grade enterprise ai systems, with a clear agent-first architecture and end-to-end workflows.

Build a simple customer support agent in Microsoft Foundry to derive a short internal summary and a polite draft reply, establishing the foundation before adding complexity.

Test the agent in the Foundry Playground with a strict customer message format, perform edge-case checks, and verify internal summaries and draft replies.

Call a Microsoft Foundry agent from a Python client in VS Code, hosted in Azure and configured via a .env file, demonstrating client-app to agent to model flow.

Tool calling in Microsoft Foundry lets an Azure-based AI agent use tools—files, databases, APIs, and internal systems—to act as a production-ready system.

Learn how Foundry configures tools in three categories—Configured, Catalog, and Custom—and how each supports enterprise use cases like file search, Azure Asearch, web search, SharePoint, and custom APIs.

Enhance a customer support agent with a file search tool grounded in the Toyota RAV4 user manual, using retrieval augmented generation, embeddings, and semantic search to answer from documents.

Demonstrates testing retrieval-augmented generation in preview by grounding answers in a Toyota RAV4 manual via file search, with citations and safety checks for edge cases.

Explore prompt engineering for enterprise AI agents using Microsoft Foundry and the Microsoft Agent Framework. Learn to create role-clear, grounded, auditable prompts that use tools safely in production.

Debug agents by tracing tool calls and file searches in the manual that build the final reply, using the traces tab to inspect conversations, reasoning, and retrieval steps.

Discover memory in the Foundry agent service, including user profile memory and chat summary memory, to deliver long-term continuity for customer support with embedding models and memory search tools.

Ground enterprise agents in external, trusted data sources with Foundry knowledge to boost accuracy and governance, using Foundry IQ and configured or catalog tools with guardrails.

Enforce strict grounding for enterprise customers by using a single source of truth, file search tools for retrieved knowledge, and I don't have that information in the available documentation fallback.

Limit what enterprise AI agents can do and say with guardrails, enforced by Azure AI Content Safety in Foundry, protecting against prompt injection.

Reassign guardrails to attach a policy to an agent, avoiding duplication and ensuring consistent behavior. Microsoft Default v2 enforces jailbreak protection, content safety, and protected materials.

Create new guardrails from scratch to define a reusable safety policy for multiple agents and models, mapping risk types, intervention points, and actions, then apply and review before assignment.

Maintain conservative defaults as a production baseline, treat guardrails as a global policy layer, separate behavior prompts from safety rules, and avoid shortcuts that weaken risk controls.

Build a simple Streamlit frontend that calls a Foundry-hosted Azure agent, enabling end-to-end testing from a minimal UI to display the agent's responses.

Test the Streamlit app with real Toyota RAV4 tickets to see the agent action end-to-end. Walk through the code at a high level, highlighting .env configuration and the UI–config–Azure separation.

Discover how AI workflows and AI agents differ in Microsoft Foundry, and learn when to use workflows, agents, or a combination for enterprise governance and decision-making.

Design and deploy your first sequential, multi-agent workflow in Microsoft Foundry. Learn how data flows between steps in a three-role marketing pipeline with a marketing analyst, copywriter, and editor.

Explore the Microsoft Foundry interface for building ai workflows, focusing on the build canvas and agent-driven data flow. Use traces, monitor, and evaluation to debug and measure production quality.

Navigate the Foundry interface to design and evolve enterprise ai pipelines, using the visualizer canvas and eml as a portable blueprint for production and versions.

Build and configure a three-agent marketing workflow in Foundry, linking a marketing analyst, copywriter, and editor. Preview the workflow and publish to a stable Azure endpoint for external Python integration.

Call a published Microsoft Foundry workflow from a Python client in VS Code to run a full pipeline with structured input and the final output.

Design and implement a reusable, multi-agent AI workflow in Microsoft Foundry, using a YouTube metadata generator with human approval loops, conditional branching, variable passing, and regeneration to power enterprise pipelines.

Create a new Microsoft Foundry workflow, save as content approval pipeline, and add a start node with an invoke agent to generate titles and a description.

Add a second agent as a refiner to the generator, forming a multi-agent pipeline that refines titles and descriptions, enables versioning, and supports humans-in-the-loop control logic.

Turn automated workflows into production-ready enterprise ai systems by implementing human-in-the-loop approvals, branching logic, variable concatenation, and regeneration loops, with structured reviewer prompts and runtime variable handling.

Learn to implement if/else logic and a regeneration loop in an enterprise orchestration pattern by routing on user approval, overwriting inputs with refinements, and looping with concatenation for iterative improvement.

Demonstrates an approval-driven workflow in Foundry that regenerates, refines, and loops with human-in-the-loop checks, delivering a final YouTube title and description and closing the production-ready feedback loop.

Discover how pretraining data pipelines convert raw internet content into a high quality 44 terabyte dataset with 15 trillion tokens via filtering, deduplication, language filters, and tokenization.

Explore how large-language models convert pre-processed internet text into tokens through byte-pair encoding, reducing sequence length and enabling learning from billions of tokens.

Train a transformer-based neural network with billions of parameters to predict the next token, updating weights via gradient descent in self-supervised learning on GPUs or TPUs.

Explore how post-training transforms a base model into a helpful assistant through supervised fine-tuning and reinforcement learning, using human feedback to shape helpful, truthful, and harmless responses.

Explore how reinforcement learning trains large language models through pre-training, supervised fine-tuning, and reward-based optimization that improves responses via prompts, feedback, and candidate solutions.

Explore how reinforcement learning enables AI to surpass human intelligence through self-play, as shown by AlphaGo's move 37 and its potential to drive AGI and ASI.

See how RLHF ranks human feedback and trains a reward model to guide language model training, compared with standard RL. Apply this to subjective tasks with no single correct answer.

Explore why hallucinations occur in LLMs, driven by vague recollection and working memory, and learn how training techniques, uncertainty signaling, and web search reduce fabrication.

Learn how stateless LLMs access current information and perform precise calculations by using internet web search, deep research, and the Python interpreter, while noting memory banks as an application feature.

Explore how large language models tokenize queries into tokens, manage a context window, and learn through pre-training, supervised fine-tuning, and reinforcement learning with human feedback.

Explore how Streamlit, an open-source Python library, lets you build and deploy web apps for machine learning and data science with just a few lines of code.

Learn Streamlit's main components and how to run a Streamlit app by importing the Streamlit library and adding charts, widgets, and tables, with caching for faster updates.

Explore how to display data with Streamlit using st.write and magic, including text, lists, dictionaries, and dataframes, powered by pandas and emoji support.

Explore Streamlit widgets by building interactive elements: text input, number input, and button; learn handling user input with st.textinput, st.number_input, st.button, and displaying results with st.write.

Show how to use checkbox, radio, and select box widgets to toggle data, preselect options with index, and display a pandas data frame based on user input.

Explore creating interactive widgets with sliders, file upload, and camera input using st.slider, st.fileuploader, and st.camerainput; handle ints, floats, dates, text, csv, and images.

Learn how to lay out a Streamlit app with a left sidebar hosting controls like a select box and a temperature slider. Drive user focus on content while adjusting options.

discover how to lay out an app with streamlit stColumns to create side-by-side widgets, control column counts and widths, and place charts, markdown, and images in multi-column layouts.

Explore the expander layout element to create a collapsible container for multiple widgets, such as bar charts and images, using with st.expander and related methods.

Display a progress bar to show lengthy computations, using st.empty as a placeholder, update text with each iteration, simulate work with time.sleep, and reveal completion with a final message.

Understand how streamlit session state stores and persists widget values across runs, enabling shared variables between sessions via st.SessionState and key-value updates.

Master Streamlit callbacks with onChange and onClick, triggering functions when widgets change, updating session state, and rerunning the script to synchronize miles and kilometers.

Explore Python concurrency ecosystem, including multiprocessing for parallelism on multi-core CPUs, threading for I/O-bound tasks with shared memory risks, and asyncio's cooperative multitasking.

Learn how async IO in python enables non-blocking concurrent code using coroutines, await, and the event loop, and compare it with synchronous equivalents.

Build an asynchronous python web scraper using aiohttp and aio-files to fetch pages and save html without blocking. Leverage coroutines, async with, and asyncio.run to manage asynchronous tasks.

Learn to run shell commands asynchronously with asyncio by creating a subprocess shell, reading stdout and stderr, and gathering results in a non-blocking, concurrent workflow.

Learn to use async ssh with asyncio to connect to a remote Linux VM, install async ssh, and run single or multiple commands with stdout, stderr, and sftp/scp support.

Run multiple SSH clients in parallel with asyncio and asyncssh to automate remote server configuration. Learn to spawn tasks, gather results, handle exceptions, and interpret exit statuses.