Contents
According to GitHub, its new architecture handled 7,200 writes per second to a single repository in testing, achieving 180 times the throughput of its predecessor. Announced at GitHub Universe 2026, the overhaul is scheduled to roll out to all users by early 2027 and is designed to support thousands of AI agents working in parallel.
Key takeaways#
- GitHub reported recording 7,200 writes per second in a test with 5,000 AI agents working in one repository.
- Branch-level locking allows writes to different branches to run in parallel, reportedly increasing throughput three- to fourfold.
- The new architecture separates compute from storage, allowing servers that handle Git requests and storage capacity to scale independently.
- GitHub plans to migrate all users to the new architecture by early 2027, but the reported benchmark does not establish that all workloads will see comparable gains.
Why GitHub is making changes#
AI agents can perform many repository operations simultaneously. When thousands of agents work in the same repository, the infrastructure needs to handle writes without making unrelated work wait.
At GitHub Universe, held October 28–29, Microsoft CoreAI Executive Vice President Jay Parikh explained the changes to GitHub’s infrastructure and Git architecture. GitHub is migrating its infrastructure to Microsoft Azure and deploying it across multiple regions. These technical initiatives are expected to be completed in early 2027.
For development teams, the important change is how many concurrent operations a single repository can accept. Increasing overall service capacity helps, but heavily used repositories also need fewer internal constraints on concurrent writes.
Reducing wait times with branch-level locking#
GitHub explained that it changed the locks used for Git writes from repository-level to branch-level locks. Under the previous approach, a repository-wide lock could make operations on separate branches wait for one another.
With branch-level locking, agents writing to different branches can proceed in parallel. According to GitHub, this change increased throughput three- to fourfold.
This makes separate working branches useful when agents execute tasks simultaneously. It does not, however, make conflicting edits compatible. Agents modifying the same code still need review and merge coordination, and operations targeting the same branch may still require sequential processing.
Scaling compute and storage independently#
The broader architectural change separates the servers that respond to Git requests from the systems that store repository data. According to GitHub, repository data is stored in Azure Blob Storage, while request servers temporarily cache data to process operations quickly.
This separation allows GitHub to reduce coordination between compute and storage and scale their capacities independently. Reference transactions that remain on the critical path are handled by a specially optimized log service. These transactions coordinate updates to references such as branches.
According to GitHub, the previous architecture could handle only a few dozen pushes per second in a single repository. The new architecture is designed for thousands per second. In the reported test, 5,000 agents working in one repository performed 7,200 writes per second, producing the headline result of a 180-fold throughput increase.
What the benchmark tells us#
The 180-fold figure describes throughput in the reported test. It does not mean every push completes 180 times faster, or that clones, fetches, code reviews, and CI runs improve by the same proportion. Throughput measures the amount of work completed over a given period, while latency measures the time each operation takes.
The explanation provided lacks the benchmark details needed to reproduce the results or predict gains for a particular team. Comparing the test with production workloads requires examining push sizes, distribution across branches, cache conditions, and write patterns. The explanation also uses both “pushes” and “writes,” which should not be treated as interchangeable without precise definitions.
As of November 4, 2026, full migration is still planned for early 2027. Teams should evaluate the overhaul against their own repository usage once it becomes available. Even if increased write capacity removes one constraint, CI queues and review capacity may continue to limit how quickly changes reach users.
Next steps#
Before migration, record baseline metrics for your busiest repositories. Over a week representative of normal work, measure peak pushes per minute and push completion times, then repeat the measurements after GitHub confirms that the repository has migrated to the new architecture.
Source: https://www.publickey1.jp/blog/26/github180aigithub_universe_2026.html
Contents
