Git hits scaling Wall as agent commits explode

The gist

Git’s human-scale code architecture is buckling as AI agents flood repos with unprecedented commit traffic, forcing major labs to reinvent core code storage for the agent era.

What to know

  • By late 2026, agent-driven commits exposed Git’s limits as merge contention and repo bottlenecks broke old workflows and consistency models.
  • OpenAI’s ten-week Pando project slashed push latency by 80% and boosted merge reliability to 99.99%, while Anthropic’s new Scion system supported 30,000+ merges daily.
  • OpenAI is buying Pierre Computer Company for its AI-optimized Git storage, boasting 500+ requests per second and support for repos up to 32TB.

Architectural Limits Exposed

Git’s distributed model, once a breakthrough for developer freedom, became a bottleneck as agent-driven commit traffic outpaced its original design for coordination and sequencing.

The late-2026 breakage looked sudden, but its roots were architectural: before Git, systems like CVS and Subversion kept repositories coherent by having a shared server assign global, monotonically increasing revision numbers, creating one authoritative sequence of states. That model made coordination legible precisely because it serialized writes through a central authority, a tradeoff that worked for human-paced development but revealed why any code-storage system built around strict global ordering would become a reliability and throughput choke point once commit traffic surged.

Git’s 2005 answer, created by Linus Torvalds in April of that year, was to “replace centralized revision numbering with a distributed directed acyclic graph (DAG) of commits identified by cryptographic hashes,” allowing multiple parents per commit and removing the need for a server to mint sequential revision IDs. But that original sequencing model was designed to free developers from centralized serialization, not to absorb late-2026 volumes of machine-generated parallel work, so the wave of agent commits exposed limits in how existing Git and code-storage implementations handled ordering, coordination, and consistency under extreme concurrency.

Concurrency Forces Backend Rethink

Surging agent commits shattered old assumptions about repository layouts and forced a shift to causal, high-throughput backends as traditional merge workflows collapsed under massive parallelism.

The first breakage showed up not as a dramatic crash but as a workflow that stopped making sense once many agents wrote at once. On Latent Space, the problem was summed up bluntly: “Merging is the wall that everyone is hitting right now,” because two agents can work “at the same time for two or three hours,” each making “30,000 line change each,” only for one to land first and force the other to redo work atop changed APIs and logging, proving that linear serialization had become the bottleneck rather than a safety mechanism.

That pressure also exposed how poorly conventional repository layouts handled agent concurrency. The same Latent Space discussion says “Jeffrey… Emanuel… we… fundamentally disagree that having 12 agents work in a single repo clone is a good idea,” with alternatives like “either get work trees with lots of branches or separate repo clones,” while another proposed a “file reservation system” because “every once in a while the file reservation gets screwed up,” underscoring that naive shared-clone Git assumptions were giving way to coordination-aware, agent-optimized repository organization.

By late 2026, OpenAI supplied the clearest proof that this was an infrastructure problem, not just a process annoyance: in a November 2 engineering post, Aaron Friel, Taylor Blau, and Ted Nyman said that starting around the end of 2025 the “old setup couldn’t keep up,” with “Fetches and pushes slowed down” and “Merges failed,” while quarterly commits landing on its internal monorepo’s master branch were projected to be “2,442% higher in the fourth quarter of 2026 than in the first quarter of 2024.” OpenAI then spent “ten weeks building Pando,” which kept Git commands unchanged while shifting to a backend that serves close to one petabyte a day, delivered more than 57 terabytes of packfiles in a peak hour with over 99% from local disk cache, and cut median push latency by nearly 80% while raising merge reliability to 99.99%.

Sources

Labs Double Down on Code Infrastructure

AI labs are racing to build or acquire code storage platforms purpose-built for agentic scale, treating version control as critical infrastructure rather than a commodity tool.

The strongest signal that this market is real is that frontier labs are no longer treating code storage as a peripheral tool choice but as core infrastructure they must own. Anthropic said, “In May 2026, Anthropic averaged around 2,500 code changes merged per day; as of September, we merged over 30,000 per day… In order to support this growth, we built Scion, an in-house version control system, and fully migrated our monorepo from git in five months,” then paired it with review, search, IAM, and landing workflows—evidence of a platform buildout, not an experiment. Anthropic also said, “We perform on the order of 25 million code checkouts per day for CI, before counting developer environments, agent sandboxes, and research jobs,” underscoring that the storage and retrieval load is already operating at platform scale.

The same pattern is showing up as a buy decision, not just a build decision. Memeburn reported OpenAI reached “an agreement to buy Pierre Computer Company… a small startup that builds Git storage designed for AI rather than people” announced “November 4, 2026,” with OpenAI saying Pierre’s “Code Storage sells ‘off-the-shelf Git infrastructure for AI applications,’” and Pierre advertising production-scale capacity of “500+ Git requests per second seen on a single repo” plus “Repos up to 32TB, with no limit on the number of repos,” a commercial footprint reinforced by named AI coding customers already using it.

Sources

Get the stories behind the trends

Deep-dive reporting and the weekly brief, in your inbox.