A Single 8-Core Server Delivers 567 Pushes Per Second: Reason Machines Open-Sources Git Server Purpose-Built for AI Agents

Reason Machines has officially open-sourced Aspen. As reported by MarkTechPost, it is a self-hosted Git server designed specifically for AI agents, following OpenAI's Pando design. The company's internal testing recorded a sustained throughput of 567 pushes per second running on a single 8-core server. The original announcement comes from MarkTechPost, written by Asif Razzaq: Link.

Git is no stranger to software developers. With core capabilities including hosting, versioning, branching, merging and rollback, it powers collaboration across human engineering teams. The dynamic shifts drastically when AI agents are integrated into the workflow: agents constantly modify files, open new branches, make commits, retry operations, and even run parallel trial-and-error tests. Git is no longer just an interface built for human collaboration, and has evolved into a underlying infrastructure for agents to write to and retrieve state from — this is exactly the niche Aspen targets.

567 pushes per second is a throughput metric. A push is not identical to a commit, but the figure clearly reflects how fast the server can receive and process reference updates. Delivering this level of performance on a single 8-core server at the very least proves that Aspen has targeted, specialized optimizations built for high-concurrency write paths. It is important to note that this test was run internally by Reason Machines, and full details around the test dataset, repository size, network conditions and persistence configurations are not disclosed in this announcement. This number demonstrates one possible upper performance limit, and cannot be directly equated to out-of-the-box production readiness.

For AI agents, the pressure on version control systems does not only come from the sheer volume of commits. As the number of parallel branches grows, so do overheads from conflicts, auditing, permission management, junk commits, and rollback operations. Standard human Git workflows operate under the assumption that every commit is intentional and passes human review. Commits generated by agents are far more similar to machine logs: extremely high in density, and packed with failures and retries. Who will handle the cleanup and governance of these commits will quickly become a new, unresolved challenge.

The core value of open-sourcing Aspen is that teams can self-host the service, keeping all code and agent trajectories on their own infrastructure. For environments that require strict auditing, isolation and cost control, this is far more controllable than fully managed third-party hosting services. Exact feature sets, license terms and deployment methods will have to be referenced from the project repository and official documentation, as the current report only covers the product positioning and performance figures, without diving into these finer details.

If you are not a coder, you can think of it this way: when AI agents run tasks, they also need a ledger that records exactly what version of work was modified at each step, and supports rolling back to previous states. Aspen is built to be the backend that powers this exact ledger.

The most notable takeaway from this announcement is that AI agent infrastructure is starting to segment into specialized, purpose-built tools. Beyond large models, classic challenges including version control, high-concurrency writing and audit tracking are now resurfacing with brand new requirements. The 567 pushes/s figure is an impressive performance benchmark, but the real-world metrics that matter next are latency, stability and operational cost when the system is used on actual production repositories.