.NET AI Community Standup: 🤖 Upgrading GenAINET to MAF v1 with SQUAD

The show opens on two launches at once: MAF v1 and a repo that now needs help

Bruno opens by celebrating the official release of Microsoft Agent Framework v1, calling out that it pulls together the best of Semantic Kernel and AutoGen on top of standards like MCP and Microsoft.Extensions.AI. But instead of doing a generic launch recap, he pivots to the real problem: the team's Generative AI for Beginners course now has to be upgraded again, and nobody wants to manually touch every project and package.

What SQUAD actually is: not one assistant, but a staffed-up dev team

Bruno explains SQUAD as a layer on top of Copilot CLI: git init, squad init, and you're off, assuming you're logged into GitHub Copilot. The hook is that it doesn't just answer prompts — it forms a team with distinct roles, and if the repo needs a new specialty, like CLI design or education strategy, it can hire that agent into the roster.

A smaller repo makes the concept click: building a Hugging Face CLI without doing the boring parts

To avoid making viewers stare at a giant course repo for an hour, Bruno switches to a small library he built for downloading Hugging Face models. He asks SQUAD to plan a .NET CLI tool with commands like download, list, info, delete, and publish as a .NET tool, and the system responds by adding a new CLI specialist named Tank to his Matrix-themed team and laying out phased work with Spectre.Console for rich terminal output.

Parallel agents, scribes, and branch safety: the workflow is the product

Once Bruno says 'implement phase one,' the fun part is watching the orchestration: background agents spin up in parallel, Scribe logs what happened, and he nudges the system to stop working directly on main and create a branch instead. The vibe is very Bruno — 'I'm too old to do this by hand' — and the pitch lands because the value isn't magic codegen, it's offloading the repetitive migration work.

Brady brings his own live experiment: a restaurant multi-agent app he forgot to upgrade

Brady jumps in with a repo he built to demo Agent Framework in a non-code setting: a restaurant service app with MCP servers for things like desserts and expo tools, traced inside Aspire and visualized with React Flow. Midstream, he realizes it's all out of date too, so he starts 'squadifying' it live, picking a pop-culture robot cast with Optimus as lead and Sunny as front-end dev, which keeps the whole thing playful instead of polished.

The real migration plan for Generative AI for Beginners: package diffs, renamed APIs, and triage by difficulty

Bruno then shows the serious version: SQUAD already analyzed the course repo, identified stable vs satellite packages, spotted namespace and client renames, and grouped samples into easier and harder migration tiers. That includes flagging older concepts like OpenAI assistants, mapping changes such as simulation clients to persistence clients, and recommending new content to reflect what MAF v1 actually supports now.

Hosted agents are the 'new sample we need' because deployment finally makes sense

The freshest concept in the stream is hosted agents: Bruno opens Azure AI Foundry to show standard agents and hosted agents side by side, with hosted ones retaining traces, monitors, and evaluations even though their model and packaging are fixed at deploy time. His pitch is simple and sticky: if local .NET agent code works, hosted agents are the cleanest path to move that same agent into the cloud without inventing a whole deployment story from scratch.

The repo is mostly upgraded already, so now humans do the review pass

By the end, Bruno flips to the new branch and shows that the course README already has MAF v1 and new hosted-agent scenarios wired in. Pablo reacts like someone who remembers the old manual grind, and that's the human payoff of the stream: SQUAD got them to the 90% mark, so their job is now reviewing lesson flow, sample names, and educational clarity instead of doing hours of repetitive migration labor.

Cost, local models, and what's next: more agentic background work

The Q&A turns practical fast: yes, SQUAD can burn tokens, so planning matters, economy mode exists, and each agent can be pinned to specific models instead of auto selection. Brady closes with a teaser that's probably the most futuristic part of the hour: a dispatch system where SQUAD watches Teams, takes a tagged request, does the work in the background, and comes back with a wall-of-text status update — which Bruno jokes he has already been on the receiving end of.