Agents are swamping Git: why OpenAI, GitHub and Meta are rebuilding version control

What this is about

If you followed the news in early November, you saw three seemingly unrelated stories. On November 2, OpenAI described a homegrown Git backend called Pando. On November 4, OpenAI acquired the startup Pierre Computer Company, and the same day Meta bought Conductor. A week earlier, GitHub had published record numbers in its Octoverse 2026 report. Put together, the picture is clear: Git — and above all the servers that host Git repositories — has become the bottleneck of AI coding. Git was built for humans who push code a few times a day. Today, agents work in loops, create thousands of branches, and trigger tests at every step.

The load in numbers

According to GitHub's Octoverse 2026, more than 40% of all pull requests in September were started by an agent. Pull requests are proposed code changes that get reviewed before they're merged. That's 110 million of them, up from a 17% share in March. GitHub itself calls the 40% a lower bound, because not every agent is detected. According to GitHub's engineering blog, there were 7.38 billion commits (saved changes) in September, more than five times as many as a year earlier. Pushes — uploads of changes to the server — rose to 3.35 billion a month, 4.9x the previous level. GitHub Actions, the service for automated tests and builds (CI), ran 3.26 billion times in September. Simon Willison calls the charts a "pretty dramatic illustration" of the impact of agentic coding.

The load had consequences. On April 9, around 84% of new sessions with the Copilot coding agent were delayed, with wait times of up to 54 minutes instead of the usual 15 to 40 seconds. The cause was a bug that applied a rate limit (a cap on requests) to all users instead of just one installation. August brought two major outages. On August 17, GitHub was disrupted for 7 hours and 47 minutes, affecting sign-in, Actions and pull requests. A core component in the US data center hadn't scaled to handle a new load peak. GitHub wrote: "Both incidents were, at their core, capacity failures" — neither was caused by a code or configuration change. During recovery, a retry loop in Copilot clients pushed the load even higher. According to Business Insider, Microsoft is also bringing in extra capacity from AWS. Microsoft confirmed this only in general terms as a "multi-cloud strategy."

Where Git breaks under agents

The engineering reports from OpenAI, GitHub and Anthropic point to four breaking points.

1. Write coordination. GitHub's existing storage system, "Spokes," typically stores each repository five times on different servers. Every push has to be confirmed by a majority of those copies via three-phase commit. Think of it as a vote: the change only counts once enough copies have agreed. A push therefore takes as long as the slowest copy needed for the majority. GitHub puts it this way: "Thousands of agents, each working on its own branch in a single repository, create a sustained write rate."
2. The branch list. On every push or fetch, the server first sends the client a list of all branches. According to OpenAI, with millions of branches this list alone can exceed 100 MiB.
3. The CI avalanche. Every push kicks off several test jobs, all of which download the same state. Anthropic counts 25 million CI checkouts a day. If every job naively fetched the whole repository, Anthropic says that would amount to more than 600 petabytes of Git traffic daily.
4. The merge queue. A merge queue lands changes into the main branch one after another, locking the entire repository to do so. Anthropic writes that such a global lock no longer works at agent speed.

Then there are the API limits. In a GitHub Community discussion from June, a developer describes how just a few review agents running in parallel trigger the so-called secondary rate limits. GitHub hasn't responded there with any substance.

Three responses: rebuild, retrofit, orchestrate

OpenAI rebuilds. According to OpenAI's blog, the systems around its monorepo (a single huge repository holding almost all of the company's code) started slowing down at the end of 2025, and merges failed. Reliability at times fell below 90%. Pando keeps the Git protocol, so humans and agents still just use git push. What's new is the server behind it. It coordinates each branch separately, so conflicts only arise when pushes hit the same branch. By default, a push to a feature branch is only told about the main branch instead of every branch. Uploaded packs are stored immutably and served through caches. OpenAI says median push time dropped by almost 80%. According to OpenAI, Pando serves just under a petabyte of Git data a day — more than 57 TB in the peak hour, over 99% of it from cache. Pando is written in Rust, has been running since April, and was built in ten weeks with help from Codex, OpenAI says. Some outlets cite a merge reliability of 99.99% after the switch. OpenAI hasn't published any code, APIs or documentation.

The Pierre acquisition fits right in. Its Code Storage service is Git hosting for machines, with SDKs and short-lived branches that agents create and throw away. Pierre lists Lovable, Bolt, Poke and Amp among its customers. According to a case study Pierre published itself, Lovable creates more than eight million repositories a week. Pierre's open-source libraries Diffs and Trees already ship in Codex and, according to The Decoder, will remain open. The deal still needs regulatory approval, and no one disclosed a price. Whether Pando and Code Storage will merge is unclear.

GitHub retrofits. According to GitHub's engineering blog post from October 6, storage and compute are being separated. The authoritative data is moving into Azure Blob Storage, an object store that handles replication and resilience on its own. In front of it sit lightweight servers that only cache data. They only need to coordinate when updating a branch pointer. Housekeeping such as garbage collection runs separately from live traffic. For the intermediate step — write coordination per branch instead of per repository — GitHub cites 3–4x more write throughput, measured in "internal tests with realistic traffic." For the full new architecture, internal benchmarks show up to 35x. According to the sources, that 35x gain has not been confirmed in production. By October 21, 99.99999% of repositories had a copy in Azure; the move to multiple regions is slated to finish in early 2027. At its Universe conference, GitHub said it had added four million CPU cores and 160 PB of fast storage in 2026. GitHub summed up the lesson itself: "Capacity alone won't get us there."

Meta orchestrates. Conductor is a Mac app that lets developers run multiple agents such as Claude Code or Codex at the same time. Each agent gets its own worktree, a separate working copy of the same repository. According to Axios, the purchase price is undisclosed; according to Runtimewire, the team is joining Meta Superintelligence Labs. What this means for users: no new sign-ups since November 4. From November 13 the cloud is read-only, and on November 20 it shuts down and data is deleted. You can pull workspaces down to your Mac with "Transfer to local." The local app, however, will only get new models and bug fixes until early 2027 and will then be phased out as well.

What's independently verified — and what isn't

The load figures and outages are well documented. GitHub records them in its own incident reports, and several companies independently describe the same problem. All the performance gains, by contrast, come from the vendors themselves. According to InfoQ, Liatrio CTO Casey Lee, for example, points out that Cursor's numbers haven't been independently verified. Also notable: Pierre still advertises Code Storage on its website with exactly the "three-phase commit with quorum" approach that GitHub is now moving away from. In the Hacker News thread on the acquisition (96 points), one user summed it up: "Agentic coding is breaking GitHub." Others pushed back, saying developer tools are a notoriously unprofitable market, and talked about "AI bros buying AI bros."

What this means for teams running their own agents

You can't get Pando or Scion, and Cursor's Origin and Entire — the mirroring service from former GitHub CEO Thomas Dohmke — are still in preview. For regular teams, that means manual work for now. A few takeaways from the reports:

• Cap the number of parallel agents. Several hands-on reports see API limits and your own review capacity, not the model, as the real bottleneck.
• Trigger CI selectively. Don't run tests on every intermediate commit of a draft, and cache dependencies. Use shallow or partial clones, which only fetch the part of the history you need.
• Throttle retries. After errors, agents should retry with growing, slightly randomized pauses (backoff with jitter). The August outage shows how retry loops can prolong an incident.
• Clean up branches. Orphaned agent branches bloat the branch list that gets sent with every request.
• Conductor users: export your data before November 20.

The bigger picture

Our take: the news from the first week of November is connected. If you sell coding agents, you now need to control the infrastructure underneath them too. Otherwise, other people's servers and their limits will slow down your product. OpenAI is picking up storage expertise and UI building blocks; Meta is getting the orchestration layer. For storage, Meta already has its own system, Sapling, which Anthropic also cites. Meanwhile, GitHub has to rebuild its platform while it's running. So far, no one outside these companies can check whether the promised speedups hold up in practice. But everyone involved agrees on the core insight: more hardware alone isn't enough. Agents need version control that no longer coordinates the entire repository, just individual pointers. The next bottleneck is already in sight — the human who still has to review all those pull requests.