Python + Agents livestream series: Resources

[pamelafox]

Join us for our 6-part live stream series on using Python with the Microsoft Agent Framework to build AI agents and agentic workflows!

Register for the series

Livestreams

Tune in for the live streams from February 24th through March 5th, or watch them after:

Topic	Video	Slides	Write-up
🛠️ Building your first agent in Python	YouTube	Slides	Write-up
🧠 Adding context and memory to agents	YouTube	Slides	Write-up
📊 Monitoring and evaluating agents	YouTube	Slides	Write-up
🔀 Building your first AI-driven workflows	YouTube	Slides	Write-up
🕸️ Orchestrating advanced multi-agent workflows	YouTube	Slides	Write-up
🙋🏽 Adding a human in the loop to agentic workflows	YouTube	Slides	Write-up

Code samples

The majority of examples will be shown from this repository:

https://github.com/Azure-Samples/python-agentframework-demos

You can either open that in GitHub Codespaces and run most examples with GitHub Models, or login to Azure and use a Microsoft Foundry model deployment.

Office hours

After every session, at 11:30AM PT, we'll have an office hours in the Foundry Discord. Bring your questions there, or post them in this discussion!

We're also recording the OH as unlisted videos on YouTube and uploading the Q&A as comments to this thread:

• Session 1 OH Recording
  • How does middleware work in the Agent Framework?
  • Why do the tools in the demos have hard-coded return values?
  • How does "context" differ across frameworks?
  • What should I do if I get an "unavailable model" error with GPT-5 Mini?
  • Is it possible to see the full information sent to the LLM?
  • Were these examples hand-coded or vibe-coded?
  • Do you recommend starting with a deployed model (Azure Foundry) for learning agents?
  • Can you use local Ollama models with the Agent Framework?
  • Are all the models you're using free?
  • Does the tracing in Agent Framework work with OpenAI tracing?
  • How does the supervisor agent pattern work?
  • Can you use GitHub Copilot models with the Agent Framework?
  • Do you always use Codespaces or only for demos?
  • What is YOLO mode in Copilot?
• Session 2 OH Recording
  • How can we configure the GitHub workspace to call paid, authenticated Azure LLMs?
  • Will this series cover hosting agents in Azure, such as options and best practices?
  • How should you manage multi-tier memory to store lots of information while keeping local storage manageable?
  • What's a good plan to learn how to build agents with the Agent Framework?
  • VS Code often forgets you have a venv in your project — does it swap that knowledge out of its memory?
  • Why would you use Agent Framework instead of the Foundry SDK, and why consider non-Foundry hosting options?
  • Can you confirm that designing your overall project cannot be LLM agnostic?
  • What are the issues with Foundry?
  • Is there any middleware for context caching that can cut down on token usage?
  • Do we have examples with App Service and UV on Azure?
  • Is it possible to use skills (like GitHub Copilot skills) in the Microsoft Agent Framework?
  • What about the TOON notation format for reducing token usage?
• Session 3 OH Recording
  • Do you need an Azure account to use the Azure AI evaluation SDK, and is there a free level?
  • Does a Microsoft 365 Family subscription include any tokens for Foundry or Azure?
  • How do you set up custom spans within traces in OpenTelemetry?
  • What if I get a deployment error because my Azure account doesn't have Owner access?
  • Why did the LLM take 5 seconds to decide on tools?
  • What about adding a validation/checker agent to improve quality?
  • Is Microsoft Agent Framework GA yet? What's recommended for production today?
  • What are the uses for observability beyond debugging and performance?
  • Should LLM-as-a-judge evaluation run in the user-facing loop or offline?
• Session 4 OH Recording
  • How do you stop an agent from outputting an old value in its context rather than re-running a function tool or an API call?
  • How do you prevent malicious prompts from proliferating your workflows?
  • What is the recommended approach for integrating the GitHub Copilot CLI and SDK in a DevOps workflow?
  • Have you tried building MAF applications using Claude Code or Codex?
  • How likely is it that a conditional workflow gets stuck in an infinite revision loop, and how do you prevent it?
  • The switch-case workflow throws an unhandled exception when running with DevUI — is it an environment issue?
  • Is there a certification about agents from Microsoft?
  • What LLMs or SLMs do you recommend for running workflows locally?
  • Where can I find all the course series?
  • What is the status of Foundry Local for Linux?
• Session 5 OH Recording
  • How do you improve the performance (latency) of a multi-agent system?
  • How do you incorporate A2A protocols into model orchestrations to integrate other agent providers into Foundry orchestration?
  • Are there limitations in workflow evaluation using Microsoft Foundry if you deploy your agent as a hosted agent?
  • What evaluation framework do you recommend — DeepEval, Ragas, others?
  • Is there a more native way to access workflow context from middleware, rather than manually injecting it?
  • How does tracing/logging work with workflows?
  • Will Microsoft Agent Framework be submitted to the AI Foundation?
  • How should you version prompts and tool descriptions for agent systems?
  • What real-world problems do these workflow patterns solve architecturally?
  • Can you use the OpenAI real-time API with Microsoft Agent Framework?
  • Can we get the full stack code for the AI finance agent?
  • Can you do breakpoint debugging with workflows in VS Code?

Related resources

If you are enjoying this series, you might also like:

• AI Agents for Beginners: Tutorials
• Python + AI: Video series (Oct 2025)
• RAG Deep Dive: Video series (January 2025)

[pamelafox]

2026/02/24: Discord OH Q&A: How does middleware work in the Agent Framework?

📹 0:01

The Agent Framework supports three types of middleware, each operating at a different level:

• Agent context middleware — runs before and after agent.run(). You get access to the agent, messages, session (chat history), and options (e.g., whether streaming is enabled). You can override the result or modify options after the agent runs.
• Function context middleware — sits between the LLM calls and the tool/function calls. Useful for security-related concerns like permission checking, human-in-the-loop approvals, limiting the number of tool calls (e.g., cutting off a deep researcher after 12 tool calls), and tool retry logic.
• Chat context middleware — operates on the chat level, where you can override or filter the LLM's response (e.g., PII reduction).

All three middleware types let you mutate the result if needed. You can define middleware using simple functions or classes.

[pamelafox]

2026/02/24: Discord OH Q&A: Why do the tools in the demos have hard-coded return values?

📹 4:01

The demo tools return hard-coded values so they work without requiring API keys. For a real implementation, you'd replace the hard-coded returns with actual API calls (e.g., requests.get() to a weather API). Most weather APIs require keys, so the demos avoid that dependency.

[pamelafox]

2026/02/24: Discord OH Q&A: How does "context" differ across frameworks?

📹 5:11

The word "context" is extremely overloaded in the AI/agent space. In the Agent Framework specifically:

• Context (as in context providers) — information that always gets passed into the agent, as opposed to tools where definitions are passed but may or may not get called. This is covered more in the session on context and memory.
• Middleware context — the context object passed to middleware, giving it access to what it needs to operate (agent context, function context, chat context).

Every framework uses "context" differently, and even within a single framework it can mean different things depending on where it appears.

[pamelafox]

2026/02/24: Discord OH Q&A: What should I do if I get an "unavailable model" error with GPT-5 Mini?

📹 6:52

GPT-5 Mini access may be more restricted for some users on GitHub Models. Workarounds:

1. Check the GitHub Marketplace models page to see if your account can access it
2. Create a .env file and set GITHUB_MODEL to a different model (e.g., gpt-4o)
3. Set the environment variable directly: GITHUB_MODEL=gpt-4o

All the examples in the repo check for a GITHUB_MODEL environment variable and fall back to a default. If deploying to Azure, GPT-5 Mini doesn't require an access request form and is available in many regions.

[pamelafox]

2026/02/24: Discord OH Q&A: Is it possible to see the full information sent to the LLM?

📹 9:52

Yes — set the logging level to debug:

import logging
logging.basicConfig(level=logging.DEBUG)

This shows the full HTTP request being sent to the chat completions endpoint, including the JSON data with the conversation, model, streaming settings, and tool definitions. Since the Agent Framework wraps the OpenAI SDK, setting debug logging will show what's sent to the LLM.

Seeing the response body is harder — the repo's AGENTS.md file has tips for how to inspect response bodies with various SDKs. Open Telemetry tracing (covered in the Thursday session) provides another way to see this information.

[pamelafox]

2026/02/24: Discord OH Q&A: Were these examples hand-coded or vibe-coded?

📹 13:54

A mix. The earlier examples shown in the session were mostly hand-coded. For later, more complex examples, the process was collaborative with GitHub Copilot:

1. Use plan mode in GitHub Copilot
2. Provide an outline and point the agent to the most similar existing examples
3. Do a lot of back-and-forth on the plan before implementing
4. Let the agent generate the code based on the agreed-upon plan

It's described as a collaborative process rather than pure "vibe coding."

[pamelafox]

2026/02/24: Discord OH Q&A: Do you recommend starting with a deployed model (Azure Foundry) for learning agents?

📹 15:53

Yes, deploying sooner is better because:

• GitHub Models rate limits are reached quickly with agent workloads since agents use a lot of tokens
• Local small models generally aren't good enough for reliable tool calling (at least on typical developer machines)
• Frontier models (GPT-5 Mini, GPT-4o, etc.) provide the best tool calling support

Even $20 worth of credits goes a long way. You can use Azure, OpenAI directly, or both. The repo's README has instructions for deploying to Azure with azd login and azd provision.

[pamelafox]

2026/02/24: Discord OH Q&A: Can you use local Ollama models with the Agent Framework?

📹 17:49

Yes, technically. The question is whether they work well. Tips:

• Use a model that supports tool calling — filter for "tools" on ollama.com models page
• Recommended models for tool calling: Qwen 3, GPT-4All (if your machine can run it), GLM models (if you have sufficient VRAM)
• When running locally, connect via http://localhost:11434/v1
• When running inside a dev container, use http://host.docker.internal:11434/v1 instead

A live demo showed Llama 3.1 successfully handling a basic agent example through Ollama.

[pamelafox]

2026/02/24: Discord OH Q&A: Are all the models you're using free?

📹 25:11

No. The cost breakdown:

• GitHub Models — free (used by default in Codespaces), but has rate limits
• Azure — not free, but used to avoid rate limits. Uses keyless connections with DefaultAzureCredential
• OpenAI — not free
• Ollama — free (runs on your local machine)

[pamelafox]

2026/02/24: Discord OH Q&A: Does the tracing in Agent Framework work with OpenAI tracing?

📹 28:00

Probably not directly. Agent Framework uses Open Telemetry for tracing, while OpenAI tracing appears to be its own thing (built specifically for the OpenAI Agents SDK). Since the Agent Framework wraps the OpenAI client, there might theoretically be a way to pass tracing info through, but it would likely not work out of the box. This topic is covered more in the Thursday session on Open Telemetry.

[pamelafox]

2026/02/24: Discord OH Q&A: How does the supervisor agent pattern work?

📹 29:06

A supervisor agent manages multiple specialist agents by wrapping them as tools:

1. The supervisor has instructions describing it manages specialist agents and should decide which to call
2. Each specialist agent is wrapped as a tool function — e.g., plan_meal is a tool that runs the meal agent with a query and returns its response
3. The supervisor can potentially call multiple specialist agents, even in parallel

Key observations from the live demo:

• Parallel tool calling can happen — OpenAI models support suggesting multiple tool calls in a single response by default
• If the agent doesn't have enough information, it may ask follow-up questions instead of completing the task. You need either a conversation loop or enough detail in the initial prompt.
• Sub-agents are also useful for reducing the context window, which will be covered in the session on context and memory.

[pamelafox]

2026/02/24: Discord OH Q&A: Can you use GitHub Copilot models with the Agent Framework?

📹 36:53

Yes. The Agent Framework has a GitHub Copilot provider:

1. Install the additional package: agent-framework-github-copilot
2. Import GitHubCopilotAgent instead of the regular Agent class
3. The Copilot CLI must be installed and logged in in the current environment

It works by wrapping the Copilot CLI binary. In the live demo, it was tricky to get working inside a dev container (required installing the Copilot CLI and logging in within the container). Once set up, you just swap Agent with GitHubCopilotAgent.

The GitHub Copilot team considers their agent runtime to be among the best available. Note that the Copilot CLI's agentic loop is actually different from VS Code's Copilot agentic loop — they implement things differently despite sharing the product name.

Links shared:

• Agent Framework GitHub Copilot samples

[pamelafox]

2026/02/24: Discord OH Q&A: Do you always use Codespaces or only for demos?

📹 42:20

Lately, more local development instead of Codespaces. The main reason is that azd login (Azure Developer CLI login) is harder in Codespaces with the current tenant setup. Working locally (still in a dev container) makes it easier to stay logged into Azure. Codespaces is still liked in general, but the Azure authentication friction has pushed more work to local dev.

[pamelafox]

2026/02/24: Discord OH Q&A: What is YOLO mode in Copilot?

📹 50:39

YOLO mode auto-approves all tool/command executions without confirmation. It's available both in the Copilot CLI and VS Code (search for "auto approve" in settings).

Caution: Even inside dev containers and Codespaces, authenticated tools (like the GitHub MCP server) can still perform real actions. The recommendation is to approve commands per session (per chat thread) rather than enabling full YOLO mode globally, since authenticated access to services like GitHub means an agent could make real changes.

[pamelafox]

2026/02/25: Discord OH Q&A: How can we configure the GitHub workspace to call paid, authenticated Azure LLMs?

📹 0:54

The python-agentframework-demos repo README has instructions for configuring model providers. If you're not using GitHub Models in a Codespace, you need to set up a .env file. Options:

• GitHub Models locally: Set up a personal access token as described in the README.
• Azure AI Foundry models: Use the Bicep provisioning included in the repo (azd up) which will write the .env file for you automatically. Or manually create a .env with your endpoint and chat deployment name.
• Other providers: OpenAI, Ollama, etc. are also supported with similar .env configuration.

The .env.sample shows the required variables: endpoint and chat deployment. By default, keyless authentication is used (no API key). If you want to use a key, you'd need to modify the code. Running azd up as suggested in the README will provision the resources and write the .env file automatically.

Links shared:

• python-agentframework-demos README - Using Azure AI Foundry Models

[pamelafox]

2026/03/03: What LLMs or SLMs do you recommend for running workflows locally?

This question was asked in Discord chat near the end of the session but the recording cut out before it could be answered on stream. Foundry Local is a tool for running models locally, available on GitHub.

[pamelafox]

2026/03/03: Where can I find all the course series?

This question was asked in Discord chat. Links shared:

• Python + AI series resources
• Python + MCP series resources
• Python + Agents series resources
• Pamela's talks page
• Pamela's Python talks slides

[pamelafox]

2026/03/03: What is the status of Foundry Local for Linux?

Currently, Foundry Local only supports Mac and Windows. Linux support is in private preview. Fill out this form

[pamelafox]

2026/03/04: How do you increase the performance (latency) of a multi-agent system?

📹 0:30

The first step is to set up quality evaluations with a ground truth baseline, so you can make performance changes and confirm quality doesn't regress. Once evaluations are in place, you can try several optimizations:

• Use smaller/faster models: Try quicker models or, if using a reasoning model like GPT-5.2 or GPT-5.3, set reasoning effort to "none" to save time.
• Reduce input tokens: Use context management techniques like sub-agents to keep context windows small. This was covered in the second session of the series.
• Reduce output tokens: Have the LLM output shorter responses (e.g., a one-sentence summary instead of a paragraph).

Always check evaluations after each change — sometimes a quality improvement causes latency to spike, making it not worth it. You need evaluations to reason about the latency/quality trade-off.

[pamelafox]

2026/03/04: How do you incorporate A2A protocols into model orchestrations to integrate other agent providers into Foundry orchestration?

📹 3:47

Microsoft Agent Framework has a sub-package specifically for A2A (Agent-to-Agent) integration. It lets you connect to an A2A agent and get responses from it. The documentation on hosting your agent with A2A protocol covers this integration. However, the specifics depend on whether you're trying to communicate with an A2A agent or host one yourself.

[pamelafox]

2026/03/04: Are there limitations in workflow evaluation using Microsoft Foundry if you deploy your agent as a hosted agent?

📹 5:48

Pamela hasn't personally deployed an agent as a hosted agent yet and couldn't speak to specific limitations. If anyone has experience with hosted agents and evaluation, they were encouraged to share. A follow-up series about hosting agents may provide the opportunity to explore this.

[pamelafox]

2026/03/04: What evaluation framework do you recommend — DeepEval, Ragas, others?

📹 6:42

For workflows, the same principles as agent evaluation apply: evaluate the final output of the workflow against ground truth, but also evaluate each individual agent along the way.

Recommended options:

• Azure AI Evaluation: What Pamela primarily uses. The advantage of using it (or OpenAI Evals with Foundry projects) is that results show up in the Foundry portal.
• OpenAI Evals: Microsoft is moving towards using OpenAI Evals with Foundry projects — worth trying.
• Agent Framework lab packages: Includes TAU2 (specifically for customer service agent benchmarks) and GAIA (for agents and workflows), though documentation for GAIA is limited.
• Custom evaluations: Remember the core principles — you're either using an LLM as a judge (outputting yes/no or 1-5 with reasoning) or some other heuristic, then doing batch evaluation. You can ask GitHub Copilot to write a custom eval framework if existing ones don't fit your needs.

For .NET developers, Agent Eval (a .NET evaluation framework) was recommended by another attendee — it includes latency and cost checking plus built-in red teaming.

Pamela also recommended subscribing to Hamel Husain's blog for everything related to LLM evaluation.

[pamelafox]

2026/03/04: Is there a more native way to access workflow context from middleware, rather than manually injecting it?

📹 10:50

This question was about middleware needing to save data to workflow shared state, but middleware doesn't have access to the workflow context. The attendee was manually injecting the context.

Pamela acknowledged this is a deep question and suggested posting it as a discussion on the Agent Framework GitHub repo, since the middleware story for workflows specifically may need improvement. The discussions and issues on the agent framework repo have been very helpful for getting answers from the team.

[pamelafox]

2026/03/04: How does tracing/logging work with workflows?

📹 12:14

It works the same way as with agents — just call configure_otel_providers() from agent framework. Pamela demonstrated this live by adding OpenTelemetry tracing to a workflow example and viewing the traces in DebUI.

The traces show parent spans for each workflow step: the agent execution, the edge (transition between agents), and each subsequent agent. So you can see the full workflow flow in the trace viewer. This also works with Aspire or App Insights. All you need is:

from agent_framework import configure_otel_providers
configure_otel_providers()

[pamelafox]

2026/03/04: Will Microsoft Agent Framework be submitted to the AI Foundation (which has MCP, Goose, and agents.md)?

📹 17:54

Pamela hasn't heard anything about this and isn't sure how projects get added to the foundation. Microsoft Agent Framework has a lot of Microsoft/Azure-specific integration, so it's unclear whether it would fit. Her observation is that protocols (A2A from Google, MCP from Anthropic) tend to come from companies developing frontier LLM models, and Microsoft doesn't have its own frontier models yet. Agent Framework tends to adopt emerging industry patterns (A2A, AGUI, MCP) rather than originating them. It would be nice if the industry agreed on common terminology, but terms like "magentic" originated from Microsoft (via AutoGen), while other frameworks like LangChain have their own orchestration concepts (e.g., Deep Agents).

[pamelafox]

2026/03/04: How should you version prompts and tool descriptions for agent systems?

📹 23:21

Since tool descriptions are in code and are part of what the LLM processes, it's hard to separate prompt versioning from code versioning. Pamela's recommendation: keep prompts in your codebase. You can put system prompts in separate files (markdown or Jinja2 templates) and pull them in, but version them alongside your code.

Tie evaluations to your PRs — Pamela showed a GitHub Actions workflow that can be triggered on PRs to run evaluations against the local app inside the runner, upload results as artifacts, and summarize them. This way, changes to prompts or tools get evaluated as part of the normal code review process.

Links shared:

• Python Agent Framework demos - GitHub Actions
• Evaluation GitHub Actions workflow

[pamelafox]

2026/03/04: What real-world problems do these workflow patterns solve architecturally?

📹 33:31

Pamela acknowledged that while she can show the patterns, she can only speak to scenarios from her own job as an advocate. She plans to automate more of her own workflows with agent-framework in the future, possibly in conjunction with the Copilot SDK for coding tasks.

Known strong use cases:

• Coding agents: The best-proven use case for LLMs — code is structured, and LLMs know coding languages extremely well.
• Fuzzy matching: LLMs can be much better than regular expressions for fuzzy matching tasks.
• New automation: Some workflows may not have existed before and are only possible now with LLM access.

She encouraged attendees using workflows and agents in production to share what works and what doesn't to help inspire others.

[pamelafox]

2026/03/04: Can you use the OpenAI real-time API with Microsoft Agent Framework?

📹 36:41

Pamela hasn't played with the newest OpenAI real-time models yet. Another advocate, Bruno Capuano, has a sample that combines real-time audio with agent framework in .NET, using whisper for speech-to-text and text-to-speech with voice activity detection. Pamela suggested reaching out to Bruno on LinkedIn for additional advice or samples, and noted that showing the overlap of agent framework with different communication modalities (WhatsApp, real-time audio) is a common request.

Links shared:

• ElBruno.Realtime - Pluggable real-time audio conversation framework for .NET
• Azure OpenAI Realtime Audio SDK samples

[pamelafox]

2026/03/04: Can we get the full stack code for the AI finance agent?

📹 40:50

Yes — the Agentic AI Investment Analysis Sample is the full-stack repo. It uses React (with React Flow), Tailwind for the frontend, and FastAPI for the backend.

Important caveat: The repo currently uses an old version of agent framework and does not pin the version in requirements, so it's hard to run right now, unless you change requirements.txt to specify the old version 1.0.0b260212. The team working on it said it should be updated within the next few weeks.

[pamelafox]

2026/03/04: Can you do breakpoint debugging with workflows in VS Code?

📹 27:34

Yes! Pamela demonstrated this live. Key tips:

• Set breakpoints on the first line of code in edge functions — not on docstrings or function signatures (putting a breakpoint on a docstring moves it to the function signature, which doesn't work well).
• For agent-level debugging, you'd need to write debugging middleware since agents are just classes without a convenient place to break.
• At a breakpoint, you can inspect the agent response, see the text output, and examine the workflow state.
• If you have OpenTelemetry console exporter enabled, you'll see OTEL spans in the debug console output as well.

Tip: You could even ask GitHub Copilot to write VS Code debug middleware for you.