- 5/29/2025
Microsoft’s annual developer conference was today and the main focus was AI agents. During the event, we got a look at multiple demos including some for Copilot upgrades, Microsoft Discovery and more. Here’s what you missed.
Category
🤖
TechTranscript
00:00Good morning and welcome to BUILD.
00:03This is the next big step forward,
00:05which is a full coding agent built right into GitHub,
00:10taking Copilot from being a pair programmer
00:14to a peer programmer.
00:16You know, here I am with all the bugs that I have
00:19or issues that I have to deal with in GitHub issues.
00:23The first one is adding a filter
00:27for user group size community page.
00:29Let's go take a look at this issue.
00:32It's nice.
00:33They say, like, I've got to go put some new filter up here.
00:37It also shows me where.
00:39I guess it needs to do a range, small, medium, large,
00:44by size or some kind of a percentile.
00:47It's like some kind of a group by case when type of thing.
00:51So anyway, so let's do the thing that is easiest for me,
00:55which is assign it to my new buddy, Copilot.
00:58So I'm going to assign it.
01:00And there you go.
01:02Let's go and see.
01:04Let me scroll down.
01:06Ah, it's picked it up.
01:11It sees me.
01:12It creates a PR.
01:13And, you know, you see that nice eye emoji.
01:17It sort of knows that I'm here and it's sort of going on to work.
01:20And we'll come back and check it out later, right?
01:22It's just so fun, right?
01:23To be able to go take care of issues like that,
01:26like email triage assigned to Copilot.
01:29It's not often that I get the chance to check on Satya's work.
01:33So let's take a look at the issue that he assigned earlier.
01:37Do you remember it?
01:38To add that group size filter to the site.
01:40Let's take a look at it.
01:42And Copilot was able to open up a pull request to start implementing these changes.
01:47Let's take a look at the PR.
01:49Copilot was able to add the group size filter property.
01:53It was able to add the logic for Redis caching.
01:56And it also added some testing with instructions.
02:00Now, it's also deployed the app to our staging environment.
02:04So let's take a look at the deployment.
02:06Fingers crossed.
02:08And you can see the group size filter was added.
02:14Today, we are introducing a new class of enterprise-grade agents you can build
02:19using models fine-tuned on your company's data, workflows, and style.
02:26We call it Copilot tuning.
02:29This app is my five-in-one hub for work.
02:32It's got chat, search, agents, notebooks, and create.
02:38And here, I can call on agents, like the researcher agent.
02:41Researcher can reason over all my work data.
02:44And with a connector, it can even tap into GitHub.
02:47So let's ask for help analyzing performance issues and prioritizing my next steps.
02:53It goes ahead and it asks me a few clarifying questions,
02:57just like if I delegated this to someone on my team.
03:00I'll give it a bit more direction, and off it goes.
03:04Now, let me show you how I built this connector in Visual Studio.
03:08I'll go ahead and add a name and a place to save it.
03:12And in seconds, the M365 agents toolkit gave me fully scaffolded code that I can modify,
03:18like creating my connection using the Microsoft Graph APIs.
03:22And then I can index any type of data.
03:25So here, I've defined a specific schema for my scenario.
03:29Next up, I'm using the GitHub API to fetch my issues.
03:33And finally, I'll ingest every issue into Microsoft Graph,
03:37so Copilot can now reason over this data.
03:40Let's see it all in action.
03:42I see that all of my issues are being indexed,
03:45and I'm getting a full log of what's happening.
03:48As the log loads, I'll scroll up and confirm that my GitHub connection was made,
03:52and that it has all the backlog items that I care about.
03:56So here, you can see that we have a specific proposal that has the content, language, and format that you'd expect,
04:03for example, from an experienced employee.
04:06Here, I've described what I want the agent to do, and I've given it some instructions.
04:10Now, I'll scroll below on the page and select which knowledge to ground this agent on,
04:16ensuring it's pulling from the right sources within my organization.
04:20We'll keep scrolling onto the tools section, and I've added this trigger that prompts my agent to begin working
04:27whenever a new RFP arrives in my inbox.
04:30So now, let's link this agent to the Dynamics MCP server to give it access to SKU and pricing data.
04:37And this will make for a more detailed proposal.
04:40So with just a few clicks, it's now connected.
04:43This server will keep the agent up to date automatically.
04:46It can also use third-party servers like DocuSign or custom MCP servers like this SAP one we created
04:53to access customer account data.
04:55So now, let's move to co-pilot tuning, which Satya just announced.
04:59And here's what it looks like in action.
05:01I'll open the contract builder agent and ask it to draft a contract using a couple of example documents.
05:08It's going to reference these documents and the custom model I created to assemble the contract.
05:14When the document is complete, I'll receive an e-bill with the draft.
05:19Let's take a look at it.
05:20I see that this contract is using our company's language, terms and conditions, structure, and format.
05:27And here's how I built it.
05:29I'll start by creating a new model.
05:31I'll give it some basic information, a name, a description, and I'll select the task type, which in this case is document generation.
05:39Here, we'll use our contract database from SharePoint.
05:43And I'll specify who should have access.
05:45So it's the contracts team and the procurement team.
05:48With that, our contract builder is now ready for data labeling.
05:52Now, once the subject matter experts have completed the data labeling process, we can complete the training for this fine-tuned model.
05:59And we're done.
06:01And from the M365 Co-Pilot app, the team can select Create Agent, select a task-specific agent, and select the contract builder model I just fine-tuned.
06:12And with that, the agent is ready.
06:15I've now scaled the work that I can deliver as a developer, empowering those closest to the business to reimagine their workflows with AI.
06:24This is Vibe Travel, a chat-based app that uses Azure AI Foundry to provide an AI travel agent to help people plan their trips.
06:33I'm planning a trip to New Zealand in about eight months, so I asked Vibe for places to ski.
06:39Hmm.
06:40This math isn't mathing.
06:42I'm pretty sure eight months from now is January 2026.
06:46We definitely need to make our AI agents smarter and give more grounded responses.
06:51We can do this by giving it files with reference data or connections to other services like Microsoft Fabric or TripAdvisor.
07:00In addition to giving it access to knowledge, we also give it access to our flight reservation API.
07:07Now, let's go back to our app and try our query again.
07:10In about eight months, it's January 2026, which is the summer season.
07:16No skiing possible.
07:18Now, let's have the agent actually book a flight on my behalf.
07:22Now, these flights look great.
07:25You just saw how Foundry made it easy to improve our agents, but there's still plenty of work to do in the actual code.
07:32Let's open Copilot Agent Mode, which is new, by the way, and I can choose between the most popular and the most powerful models from all the major LLM providers.
07:43Agent Mode takes GitHub Copilots beyond just answering questions or suggesting code to actually helping me do the work.
07:52I can come in here and I can ask, can you show me the issues that are assigned to me in this repository?
07:58Traditionally, I would have opened the browser to look at the issues on GitHub.com, but with MCP or Model Context Protocol, this allows you to give tools like GitHub Copilot the context and capabilities it needs to help you.
08:13Now, when I click this reset button in the app, it immediately clears our chat session without a warning.
08:20That's not ideal.
08:22And I can ask GitHub Copilot to help me implement the details for issue number one and also fetch the details for issue number one.
08:30I also have this beautiful design.
08:33It's pretty sweet.
08:34Wouldn't you say?
08:37Let me add this to the context window and send the prompt.
08:40You can see the GitHub MCP server fetch the details, and with Agent Mode's new vision capabilities, it means that Copilot can even understand the sketches of what I want.
08:50Copilot Agent Mode was able to examine the proposed changes, look for relevant files to change, make the changes as appropriate, and it even stuck to the styles and coding standards that I wanted it to.
09:03When I go back to the app, it was able to implement the UI changes that my boss wanted.
09:09Let me see if it's running.
09:11As you can see, folks, this is live.
09:23I don't have time to debug, but once I click this issue, I'm pretty sure it's implemented.
09:29Copilot was able to add the modal in our app, and that's two issues that we worked on in a handful of minutes.
09:36Thank you, GitHub Copilot and Azure AI Foundry.
09:40But there's one more domain that I want to talk about to close out, which is science.
09:46We'll have real breakthroughs in the scientific process itself, which will really accelerate our ability to create a new material, a new compound, a new molecule.
09:59That's our ambition with Microsoft Discovery, which we are announcing today.
10:05I'll show you how I lead a team of agents to make a discovery for immersion coolants.
10:10These are an interesting line of research for cooling data centers.
10:13Unfortunately, most of them are based on PFAS, or forever chemicals, which are harmful for the environment.
10:19My first step is to conduct research.
10:22I want to be up to date with the latest knowledge on this topic.
10:26I'll start by asking about coolants and their properties to help me identify starting candidates.
10:32What we see here is a summary on the left and a comprehensive report on the right covering the state-of-the-art in coolant research with links throughout to citations for trusted research.
10:45Now, I can validate these findings and iterate if I choose before moving on, but we want to make a real coolant discovery.
10:53So we need to move on to step two, which is generating hypotheses.
10:58I am going to ask for a plan specific to my investigation informed by this research, knowing, for example, that I should target a specific boiling point and a dielectric constant that won't fry my electronics.
11:13Now, you see, I didn't specify what methods to use or write any code.
11:19The agents can use tools and models from Microsoft.
11:22They can integrate open source or third-party solutions, or even ones from my own organization.
11:27And if we look right here, we can see the plan that my agents returned.
11:31It starts with a generative chemistry step here, where it creates millions of novel candidates that are more likely to meet my criteria.
11:38And then it uses AI models to screen these down quickly.
11:43And finally, we use HPC simulations for validation of our findings.
11:47We click proceed right there.
11:49And with that, we're in step three, experimentation.
11:52We can see here now the discovery agents working together in real time, driving all of these intense computations.
11:59With traditional methods, getting to a shortlist of candidates could take months or years of trial and error.
12:06Microsoft Discovery can compress the time to days or even hours.
12:11So let me show you the final results of the process.
12:14These are the set of candidates that Microsoft Discovery has identified for PFAS-free immersion coolants.
12:21But I know what you're wondering.
12:22Did we actually make a discovery?
12:25Well, this is not just a demo.
12:28We really did this.
12:30We took one of the most promising candidates and synthesized it.
12:34They didn't let me bring a new material unknown to humans onto this stage.
12:39But I've got this video from the lab.
12:42So we can see there my coolant.
12:44And we dropped a standard PC in it.
12:46And it's running Forza Motorsport.
12:49And it is keeping the temperature stable with no fans.
12:52It's literally very cool.
12:57We found a promising candidate for an immersion coolant that does not rely on forever chemicals.
13:03Imagine using Microsoft Discovery across domains, designing new therapeutics, new semiconductors, or new materials.
13:11It worked for us.
13:14The next great breakthrough is yours to discover.
13:18Have a great build.
Recommended
15:04
|
Up next
1:06:15
27:47
1:04
6:15
8:12
1:38:37
1:45:43
1:22:27
11:22