What the hell is going on with GitHub?

Summary

GitHub COO Kyle Daigle says the platform is facing unusually rapid growth as developers use agents to make and push far more changes. GitHub now handles roughly 7–8 billion commits a month, up from about 1–2 billion a month earlier in the year. Contributions this year exceed those of the previous two years combined.

Because agents behave differently from human users, they create new infrastructure challenges. Daigle says GitHub’s uptime has improved but is not yet perfect.

Infrastructure and CI

• GitHub is adding about 4 million CPU cores, largely to support GitHub Actions, which Daigle says handles around 45 billion runs.
• A larger infrastructure change separates code storage from compute. GitHub plans to store repository data in blob storage and take a more optimistic approach to updating refs, rather than waiting for a quorum of servers before confirming a push. The goal is to reduce locking and latency, and make operations such as shallow clones more efficient.
• GitHub is moving more infrastructure to Azure across multiple regions. Daigle expects the migration to deliver its full benefits around January–March. Some especially busy, large monorepos require more work because they are particularly sensitive to latency.
• The growth of agent-driven development may make CI and testing a new bottleneck. Daigle expects teams to reconsider which tests they run, potentially using more selective testing or proof-driven verification rather than simply adding more compute.

Pull Requests and Agent Workflows

The speakers discuss whether pull requests are becoming less useful when maintainers receive AI-generated contributions that take time to review. Daigle sees PRs continuing to serve as records of what changed and why, especially for teams.

For open source, he suggests a different possibility: contributors could provide a feature request and “tokens” or compute for maintainers to use in their own agent workflow, rather than submitting code that maintainers must redo or verify.

GitHub also wants to make Git easier for newcomers and agent-heavy workflows. Daigle describes a goal of progressively revealing the Git features a user needs while handling more complexity behind the scenes.

Copilot and Copilot Fusion

Daigle says he would like Copilot to let individuals bring existing AI subscriptions, though this presents technical and product trade-offs. GitHub’s Copilot API routes requests across providers to manage latency and errors, while subscription-based access may require direct provider connections.

He describes Copilot Fusion as more than a conventional model router. It can use:

• Single-model routing for straightforward tasks.
• Cascade mode, which starts with a lower-cost model and escalates to stronger models if its draft does not meet the task’s intent.
• Critique mode, in which another model reviews an implementation and helps refine it.

Daigle says Fusion is available through Copilot experiences including the Copilot app, VS Code, CLI, GitHub.com, and cloud agents. He claims its multi-model approaches can be up to 77% cheaper, are typically 50–60% cheaper, and are around 40% faster than the comparison approach.

npm and Supply-Chain Security

Daigle says GitHub is working on npm security, including:

• Unpacking packages in virtual machines to scan for malware.
• Examining dependency trees to determine whether packages’ code is actually used.
• Offering a package firewall that can restrict use to trusted packages or packages published for a minimum period of time.

He notes that AI can assist both defenders and attackers.

Daigle pushes back on the idea that developers should simply generate their own replacements for dependencies. In many cases, using and securing an established package is more practical than recreating common functionality. He also says GitHub has staff focused specifically on npm.

Code Storage and the Future of Git

Asked about newer code-storage architectures, Daigle says GitHub’s blob-storage direction shares some broad ideas with other approaches. He says GitHub will keep improving Git rather than replace it unless an alternative is materially better, given the compatibility impact of changing a foundational developer tool.

Other Mentions

• Sentry is promoted in a sponsor segment as a tool for monitoring application errors.
• For his “sick pick,” Daigle recommends a leather-bound notebook, which he uses to collect thoughts before sharing them with an agent.
• The discussion is framed around GitHub Universe and the importance of celebrating open-source projects and the people behind them.

Reviews, Guides, or Tutorials

None are presented; this is an interview and analysis discussion.

Main Speakers and Sources

Scott and Wes of Syntax; Kyle Daigle, GitHub COO. Sentry is mentioned in a sponsor segment.