GitHub COO Kyle Daigle on agents, uptime, and company workflows

GitHub is adjusting to a wave of AI-generated code and agent-driven workflows that is pushing its infrastructure harder than the systems were originally built to handle, according to GitHub COO Kyle Daigle. In a conversation tied to Microsoft Build and the AI Engineer World’s Fair, Daigle said the company is now dealing with scale pressures that come from code being produced faster, by more people, and in larger volumes than before.

Daigle said the shift is not just about software output. It also changes how GitHub runs its own business, how open source contributions are reviewed, and how teams communicate internally when AI tools are involved in daily work.

He described his own role as having expanded beyond GitHub operations into what he called a CMO of developer role for Microsoft. Daigle said that after years focused on leadership, he has been coding more again because of AI tools, and that much of his use of agents is about connecting different data sources rather than simply writing new software.

One example he gave was a retrospective workflow that pulls together information from GitHub, Teams, Slack, email, Obsidian notes, and meeting transcripts. Daigle said he uses that context to ask what happened over the past week, what worked, what did not, and what should change over the next few days.

He said GitHub has tried to roll out AI internally without forcing employees to change how they work. Instead of introducing a brand-new system for everyone, the company has been building what he called skills, giving people access to a CLI, and connecting AI tools to the services they already use.

Those services include GitHub, Teams, email, and Slack. Daigle said GitHub uses Teams for video and Slack for chat and command-based workflows, and that moving off Slack would be expensive because so much of the company’s tooling is built around it.

He also said GitHub has been using Microsoft’s WorkIQ MCP server to access company context across Microsoft 365 data. MCP stands for Model Context Protocol, a standard for connecting AI systems to external tools and data sources.

Daigle said the company’s internal AI approach is increasingly based on “micro-skills” rather than large, all-purpose workflows. He said the earlier idea of a single skill that could produce a full report or manage a broad process has given way to smaller pieces that do one thing well, because large workflows become brittle as information changes.

The same pattern, he said, is showing up in how GitHub thinks about agents, GitHub Actions, and company automation more broadly. Daigle said GitHub is also using its own tools, including the new GitHub Copilot desktop app, to run workflows every time he opens his laptop.

The discussion also returned to GitHub’s longer history, including webhooks, APIs, Actions, npm, Dependabot, and Semmle. Daigle said those products helped shape GitHub into a platform for code collaboration and automation, but the AI era is putting new pressure on its systems.

He said GitHub Actions has become a general-purpose compute layer for automation and agentic workflows, while security questions around npm, token invalidation, 2FA, and dependency trust remain central as more AI-generated code reaches repositories.

Daigle also discussed how pull requests may change when more of them come from agents. He pointed to ideas such as prompt requests, vouching, and AI review, suggesting that trust and review processes will need to adapt as more code is produced by software rather than people.

GitHub’s broader position, Daigle said, is that the company wants to keep exposing the code rather than hide it, even as low-code tools like GitHub Spark lower the barrier to building. The challenge, he said, is keeping GitHub usable for both human developers and the growing number of agents now operating inside software teams.