How GitHub is building the multiplayer future of code with Luis Bitencourt-Emilio

• Why GitHub bets on model choice over a single 'best' model, and how fleets of agents mirror what makes human teams effective

• The six-month shift that took AI coding from advanced autocomplete to something that 'just worked'

• Why human-curated data from Reddit and Wikipedia became more valuable, not less, in the age of generated content

Let's dive in. No floaties needed…

Revenge of the Nerds

Luis Bitencourt-Emilio, SVP of Copilot Agents Product at GitHub

Luis Bitencourt-Emilio is the SVP of Copilot Agents Product at GitHub, where he leads the work of turning the platform into the place where humans and AI agents collaborate to build software. Born and raised in Brazil, where his father enrolled him in a programming course at age eight on a 286 in the family's laundry room, Luis went on to a two-decade career spanning Microsoft, Reddit, and two Latin American unicorns. At Microsoft, he spent close to a decade shipping major products; at Reddit, he founded and led the company's first AI and machine learning team; and as a C-level executive at Rappi and Loft, he ran product and technology for some of Latin America's most valuable startups. He holds Brazilian, American, and Italian citizenship, is fluent in English, Portuguese, and Spanish, and is the father of a young son.

Beyond his day job, Luis serves on the Board of Trustees of the Wikimedia Foundation, which oversees Wikipedia and other free knowledge projects. He joins us at a pivotal moment. GitHub Copilot is marking its fifth anniversary, tied to the technical preview that launched in 2021, and GitHub has been shipping agent-native features at a rapid clip, from the Copilot desktop app to its first open-weights model.

With GitHub now home to more than 210 million developers and over 630 million repositories, Luis is helping steer how the world's largest developer platform adapts to an era where agents write, review, and ship code alongside people.

There are powerful coding tools out there. What does Copilot have that they don't?

The way we see GitHub is as a platform. I love all those tools, I've used all of them, and we use Copilot too. In general, they're all running on GitHub anyway, and that's the key thing we see for our developers: developer choice.

A big differentiator is that it's multi-model. Copilot lets you get models from all the top labs, and we just released our first open-weights model. The other key thing is the GitHub graph. A lot of the context and future intelligence we'll have is built off your issues, your PRs, your source code, and the conversations around them. When you think about how much data you have in your repos, it's a lot more than just source code, and having that ingrained in everything we build and train is a huge differentiator.

Then there's the Microsoft ecosystem. GitHub isn't just GitHub. When you look at the integrations with Azure, with Foundry, our COO Kyle Daigle recently published the Developer Manifesto, and there was a piece about coming at customers with a holistic view of what makes sense for the developer, versus separate products they pick and choose. So those three pieces: multi-model, the GitHub graph, and the Microsoft ecosystem. But the key is still that our goal is to be a platform that gives developers whatever they need to get the work done.

Walk us through the multi-model approach. How does the Microsoft model fit in?

I've managed teams for a couple of decades, and one thing we've learned about teams of humans is how critical diversity is. You don't hire a bunch of people great at exactly the same things. The most effective teams are the most diverse ones that can move fast. As LLMs bring technology closer to human behavior, those same paradigms start to transfer to teams of agents. Would I always use the exact same model for every task? Absolutely not. Some are great at PRDs and emails, some at coding, some at finding bugs. We see this every single day.

So the key is providing fleets of models, sometimes collaborating with each other. We have the rubber duck feature where different models discuss an implementation together. The Microsoft models build on that developer choice. We want every major lab's models there, open-weight models there, and our own latest and greatest from the Microsoft AI collaborations. When you look at the ROI, the token cost versus the outcomes, there's a lot of variance between model families, which makes it critical for developers to build a taste for all the tool sets.

But that's complex. Imagine I tell you to go learn five or six model families, then within each figure out the cheaper ones, the thinking models. That's a lot to onboard onto. So we're putting significant effort into an auto mode that picks the best model for a challenge based on the task and intent.

You mentioned running a fleet of agents in parallel. How does that change the speed at which software gets shipped?

Immensely. I got into tech early, on a 286 in my parents' laundry room in Brazil, and back then all that mattered was how many megahertz your CPU had. That world ended a while ago. We started talking about cores, then GPUs, parallelizing more workflows, because doing things serially takes a long time. These models take time too, so a big decision you make as a developer is how quick versus how deep an answer you need.

Parallelization works like talking to a human. Here's my goal, here's a GitHub issue, please solve it. Some of it needs a plan first, so you use a particular model for that. But once you have the plan, there's a bunch you can build in parallel, just like a team of engineers subdividing a task. What excites me more is the collaboration. Diverse teams are effective, but only if they collaborate. So when one model generates code, and a different model has the tooling to be critical of it- linters, automated reviews, cyclomatic complexity, license checks- you get outputs much closer to what a human team would build.

What's different today about building AI products versus two years ago?

It's vastly different versus even six months ago. GitHub Copilot just hit its fifth anniversary, and next week I'm speaking at We Are Developers in Berlin, where the first public demo happened years ago. Back then, Copilot was a very advanced autocomplete. But six months ago, around last November, the world changed. I took time off over the holidays and used the latest models, and I don't know that they were an order of magnitude better, but they crossed that Rubicon where you suddenly see the value.

We'd shipped the Copilot CLI in September, and I wanted it to help write my mid-year reviews by looking at my Slack and GitHub. I started manually piping through MCP configuration changes, then halfway through thought, what am I doing? I'm the VP of Product for GitHub Copilot; I should be doing this with Copilot. I kicked off a job, it ran once, and it just worked. Then I spent the whole holiday coding, finding where the limits were. They still exist, but every new model and this competition between the labs has driven the advancement immensely fast.

Let's acknowledge it's changing fast in my bubble. Globally, it's not changing that much yet. Pick the average S&P 500 or non-tech company, and the impact is far more gradual, but that's a matter of time. You saw the Microsoft Frontier announcement, which is about disseminating this technology and empowering every company. What we're seeing in our teams is a full-stack builder mentality, fewer hops between idea and production. When I founded the Reddit Intelligence Group, I saw teams where data scientists handed models to an ML team that handed code to an engineering team. I said I just want one team, because every hop is an opportunity for misunderstandings. You don't have to be the best at everything, but a bit of a jack of all trades who can go end to end. Now agents help with that. The parts where you're just okay, you can outsource to agents. I'm not great at creating design, but I have taste for it, and now models can work that front for me.

Walk us through the recent launch of the Copilot app.

We recently made the GitHub Copilot app more public. Folks still swear by Vim and Emacs, and the CLI had a resurgence, which I love. But there's a lot a graphical interface can do that CLIs struggle with. Our mission is to raise the roofs and lower the floors, to help more developers enter and be effective.

We recently shipped canvases, a way to build UI that both you, in your chat, and your agent can interact with together. It's rethinking human-computer interaction in the age of agents. Does this replace the CLI? Not necessarily. If I'm automating workflows, the CLI might be my go-to. But for visual, interactive experiences with humans and agents alike, the app is on the bleeding edge. You can build workflows where every morning it finds high-priority issues, sets off agents, and gets me a PR to review, so I show up to a list of things I need to handle while my agents take care of the rest.

You worked at Reddit and serve on Wikipedia's board, two of the most cited AI training sources. Are they good sources for AI training?

When I was at Reddit, Wikipedia was the source of the first knowledge graph we built. We did vector math to see subreddit similarities, even subreddit arithmetic. A subreddit big on women's topics plus one on sports gave you women's sports. We leveraged Wikipedia for that graph, then retrained it on the Reddit ecosystem. Both are unique and special communities.

The value of Wikidata is huge. A source of trustworthy, unbiased information that brings knowledge to underserved communities who couldn't buy hundreds of dollars of encyclopedias, that value is undeniable. The value of human conversations online was always clear to me, but it's even more valuable now. Reddit just ran its first consumer commercials in about five years, all around the value of human conversations.

As more generated content comes online, the signal-to-noise ratio gets thrown off. People are acquiring pre-2024 code bases to train their bots, because artisanal handwritten code now has higher value than generated code. These training systems are stochastic, so if the information is bad, it magnifies the bad outcome. That's why human-curated data matters more. Some of the fastest-growing companies of the past couple of years label data sets with expert humans, like X-ray data labeled by doctors. We also see big models used to train smaller ones, but there's a strong limit there: quality gets worse the more generated data you use. Today, gold-tier data is artisanal, human, and curated. Reddit is still among the top results on search engines, because when a strong community upvotes good content and downvotes bad, the signal you can train on is huge.

If Copilot gets this right, what does a software engineering team look like in five years?

I'd broaden that to the industry. If we get this right, you'll have more empowered teams and individuals. One person doesn't have to be an expert in everything to start a company or solve a problem. I've been on a workout kick and never found the app I love, so I started building one. It can be perfect for me, and I don't have to worry about other users. Solving challenges for yourself, then figuring out which ones become businesses, is its own entrepreneurship kick.

It's agency-heavy. You see a challenge, you tackle it, and you have all the tools to do it. You can chat with these models about how to incorporate a company, things you wouldn't have learned in school, have them help design and code your product, find bugs, get it to production, and monitor it so you only get paged when something's really serious. So it's highly empowered, autonomous teams working full-stack on these challenges.

Additional Reads

• The Blueprint for Agentic Development

• WWC26 - Craftsmanship in the Age of Agents

• Luis Bitencourt-Emilio Joins the Wikimedia Foundation Board

Quick Poll

Rate This Edition