Overview of Syntax 1055: What the Hell Is Going on with GitHub?
In this episode of Syntax, hosts Wes Bos and Scott Tolinski talk with GitHub COO Kyle Daigle about outages, the surge in agent-driven development, and how GitHub is adapting its infrastructure and products. Daigle says GitHub is seeing unprecedented growth and is redesigning parts of its systems to handle agents’ faster, less human-like activity.
GitHub’s Growth and Infrastructure Changes
- GitHub’s commit volume has climbed from roughly 1–2 billion per month in early March to 7–8 billion per month. Daigle says contributions this year have already exceeded the previous two years combined.
- Agents are driving much of the increase: a developer might once have pushed a handful of times in a weekend, while agents can push many times more.
- GitHub is adding 4 million cores, mostly for Actions, which handles about 45 billion runs.
- GitHub is separating Git’s compute and storage. Code will live in blob storage, while refs can be updated optimistically and become consistent over time. This should reduce locking and latency, support independent work across branches, and make shallow or selective checkouts more efficient.
- The cloud migration was planned for January through March. Large, busy monorepos are among the hardest repositories to move and may take additional work.
- Daigle acknowledged that uptime still needs improvement and said GitHub is continuing to address how agents behave differently from human users.
Testing and Pull Requests in an Agentic Workflow
- More agents mean more CI runs, but simply adding compute can make testing prohibitively expensive. Daigle expects teams to reconsider which tests they run and when.
- Possible approaches include more selective testing and “proof-driven development,” where teams verify a change through computed proofs instead of running every test.
- Daigle sees pull requests as useful records of what changed and why, especially for teams and closed-group projects. For open source, he suggested a model where contributors could fund or supply compute for a requested feature, letting maintainers decide whether to pursue the idea without taking on as much review work.
Copilot and Multi-Model Workflows
- Daigle would like individual developers to be able to bring subscriptions from other AI providers into the Copilot app. GitHub is considering the technical and user-experience tradeoffs; users can currently bring their own API key.
- Copilot Fusion can use multiple models for a single task. Its “cascade” mode starts with a less expensive model and escalates when needed; “critique” uses another model to review and improve the work.
- Daigle said these approaches can be up to 77% cheaper, and are typically 50–60% cheaper and 40% faster for GitHub and its customers. Fusion is available wherever the GitHub Copilot harness is supported, including VS Code, the CLI, GitHub.com, and cloud agents.
- GitHub is also exploring ways to make source control easier for people new to Git by revealing only the features they need and handling more of the complexity behind the scenes.
Package Security and Dependencies
- GitHub is scanning packages by unpacking them in virtual machines to look for potential malware. It is also improving visibility into dependency trees and developing package-firewall controls, such as allowing only trusted packages or packages that have been available for a minimum period.
- Daigle noted that AI can help defenders find vulnerabilities, but can also help attackers.
- He cautioned against automatically replacing dependencies with generated code. For many common tasks, using a well-maintained package remains more practical—and may mean fewer pieces of custom code to secure.
- Daigle said GitHub has dedicated staff working on npm and package security, alongside work across Microsoft’s package-registry services.
Git, Code Storage, and Developer Priorities
- GitHub is exploring more efficient code-storage designs while remaining committed to improving Git until a replacement offers a material improvement. Daigle emphasized that changing Git’s API would affect developers and services across the internet.
- The episode closes on a people-focused note: Daigle said GitHub wants to support developers’ ambitions and celebrate what they build, even as the industry concentrates on models, benchmarks, and tools. His personal recommendation was a leather-bound notebook for capturing thoughts before sharing them with an agent.