My Top AI Sessions at Microsoft Ignite 2026
The AI sessions I'd pick at Microsoft Ignite 2026 as a developer building agents, from first prototype to production, plus my own talks and tips for avoiding clashes.
Microsoft Ignite takes place November 17 to 20 in San Francisco, and the session catalog is enormous. If you open it without a plan, you'll spend more time scrolling than watching.
So here's my plan. These are the AI sessions I'd pick as a developer who builds with agents, from the first prototype to running them in production. I work at Microsoft, so take that into account, but these are the sessions I'd recommend to a friend.
Good to know before you start: the digital pass is free, keynotes and breakouts are streamed and recorded, and most recordings appear within a day or two. Labs, lightning talks and table talks are not recorded, so those only work in person. All times below are Pacific Time (PT). The codes, like BRK240, are the session codes you can search for in the session catalog.
My own sessions
I'm speaking too, so let me start there, and then I'll get out of the way.
Lost in San Francisco: Build AI agents that find their way
How Microsoft Foundry, GitHub Copilot, VS Code, Microsoft Agent Framework and MCP (Model Context Protocol) form one connected loop for building agents that are grounded and observable. Less slides, more building.
Governance of tools and skills with Foundry Toolboxes
A fifteen-minute lightning talk on what happens when you don't have one agent but a whole ecosystem of agents, tools and skills, and how reusable toolboxes keep that maintainable. Lightning talks aren't recorded, so come by if you're there.
Building multi-agent systems. Lost in San Francisco
The hands-on version: build a multi-agent system yourself. Labs aren't recorded either, so this one is for the people in the room.
Start here
Microsoft Ignite opening keynote
The keynote sets the direction for everything else announced that week. Even if you only watch the first half hour, you'll understand the rest of the sessions better.
Microsoft's own models
If you followed my series on Microsoft's AI models, these two are the sequel.
Turn enterprise data into a durable advantage with Microsoft AI models
A first look at the latest MAI models for reasoning, coding, voice, image and multimodal work, and why it matters that they're built on clean, traceable data.
Building a model for software development from the ground up
How Microsoft AI built a model for coding and agentic software development, with GitHub Copilot as its proving ground. The description starts with my favorite line of the whole catalog: benchmarks alone can't predict how a model will perform in a real product.
Building agents
Build and scale the next generation of AI agents in Microsoft Foundry
The core Foundry session for developers: from local development to hosted agents, Toolboxes and Foundry IQ, and the path to production. If you watch one session about building agents, make it this one.
How much should you own? Choosing the right agent architecture
How to pick the right build path for each agent, from a simple prompt agent to a managed harness, without building yourself into a dead end. It's the "can you leave?" question from my series, applied to agent architecture.
Building agentic retrieval: From DIY to Foundry IQ
First a retrieval loop built by hand, then a comparison with what a managed knowledge base does for you. You see what's happening under the hood before you hand it over, which is exactly how you decide whether to build or buy.
Voice agents in Microsoft Foundry: Prototype to production in 45 mins
Voice agents across web, mobile, Teams and telephony. If you read my series on Microsoft's AI models, this is the help desk from Day 5, taken all the way to production.
Agents in production
This is where most teams struggle, and where I'd spend the most time.
From traces to better agents: The Foundry improvement loop
How to turn what your agents actually did into evaluations, and those evaluations into improvements. A repeatable loop instead of guessing.
Token economics: Get more from your production AI budget
Matching workloads to models, reusing prompt caches, routing requests and handling capacity limits. The cost side of agents, which decides whether a pilot survives contact with real traffic.
Agents in production: Observe, evaluate, and optimize with Foundry
Monitoring quality, reliability, speed, safety and cost across many agents at once. I've written that you can't manage a team of agents you can't see. This session is about seeing them.
Run secure agent loops in production with isolated sandboxes at scale
Running agents that execute code safely, in isolated sandboxes. The fence around the agent, at scale.
Trust and governance
Operate AI at enterprise scale with Microsoft Foundry
Governance, agent identity, Responsible AI controls and the cost side of AI. Less exciting than a live demo, but this is what decides whether agents ever leave the pilot phase.
Manage and secure your AI agents at enterprise scale with Agent 365
How organizations keep track of all their agents: who owns them, what they can access, and what happens when one misbehaves.
Running AI where you are
Build private, responsive on-device AI agents with Foundry
Building agents that run on the device itself, across Windows, macOS, Linux and mobile. Private, fast, and still working when the Wi-Fi doesn't.
From open weights to workload: Operationalizing at enterprise scale
Open-weight models give you more choice and control, but running them in production takes the right infrastructure. If you read my post about owning your weights, this is the practical side of it.
Build sovereign AI with Azure Local and Foundry Local
Models, agents and retrieval in datacenters, factories and branches that the customer controls. For everyone whose data can't leave the building, or the country.
For the builders
Ship more software with parallel agents in the GitHub Copilot app
Running several coding agents side by side and keeping track of their work.
Hack the harness: Copilot SDK agents from prototype to production
Going deeper into the tools, MCP (Model Context Protocol) servers and hooks behind agents built on the Copilot SDK. For developers who want to shape their agents, not just use them.
Scott and Mark learn to... live again
Live, showing how far AI-assisted development has come. Always fun, and they tend to show what breaks, not just what works.
Meet the community
Foundry Discord Live: Product Updates & Community AMA
The latest Foundry updates and your questions answered live, with Nick Brady, Amy Boyd and Lee Stott. Sessions like this are where you can ask the question that isn't on any slide.
If you're there in person
Continuous improvement for agents: From open source to Foundry
A hands-on lab on measuring and improving an agent step by step. Labs aren't recorded, so this one is for the people in San Francisco.
Three tips for a better Ignite
Watch for clashes. Some of the best sessions run at the same time:
- Tuesday 2:00 PM: my own BRK555 and BRK240. Watch one, then the other on demand. I won't be offended.
- Tuesday 3:15 PM: BRK279 and BRK270.
- Wednesday 11:30 AM: BRK249 and BRK247.
- Wednesday 3:00 PM: BRK251 and BRK277.
- Wednesday 4:15 PM: BRK448 and BRK214.
- Thursday 2:45 PM: BRK300, BRK253 and BRK543, all three at once.
- Thursday 4:00 PM: BRK244, with the community session starting at 4:20.
Pick one live, catch the others as recordings.
Pick two or three sessions a day, not ten. You'll remember more and actually try what you saw.
Try one thing the same week. The best way to get value from a conference is to build something small with what you learned while it's still fresh. It's the same advice I gave in You Can't Keep Up With AI: small experiments beat endless reading.
See you in the stream, or in the hallway.
Pass it on