When Agents Scale, the Plumbing Breaks First
When you give agents real work, the constraint does not disappear. It moves. It moves from the people and the agents doing the work to the shared plumbing underneath them: the queues, the review steps, the systems that hold everything together. Two of the most automation-heavy companies on earth just proved it in public, with the same bottleneck, three weeks apart.
Two labs, one bottleneck, three weeks apart
On October 13, 2026, Anthropic published the story of Scion, the version control system it built to replace Git for its own monorepo (Anthropic, "Redesigning source control at Anthropic"). On November 2, 2026, OpenAI published the story of Pando, a new backend for Git repositories that now powers all development in its monorepo (OpenAI, "Scaling Git at OpenAI").
The trigger was the same in both cases: agents writing code at a volume human-scale infrastructure was never designed to carry.
- Anthropic merged around 2,500 code changes per day in May 2026. By September, it merged over 30,000 per day (Anthropic, published October 13, 2026).
- OpenAI described concurrent pushes, merges, fetches, and clones driven by Codex overwhelming its monorepo systems, with reliability "at times falling below 90%" around the end of last year (OpenAI, published November 2, 2026).
These are extreme workloads. But the direction applies to any business running agents at real volume. Anthropic put it plainly: "the latent demand for merging code is practically infinite, and every bottleneck we remove immediately leads to a further increase in the amount of code we produce." Swap "merging code" for whatever your agents produce, and the sentence still holds.
What actually broke
The details differ, but both posts describe the same shape of failure. The agents were fine. The shared layer around them was not.
At Anthropic, the monorepo's main branch history alone is almost 100 GB, and the company performs on the order of 25 million code checkouts per day for CI. Fetched naively, that would mean over 600 PB of Git traffic per day. The stopgap was running several hundred internal Git mirrors at peak, which became harder to operate as the merge rate grew (Anthropic).
At OpenAI, the problem was contention. More engineers, and more agents, pushing to more branches at the same time meant slow fetches, slow pushes, and failed merges. Its fix, Pando, cut median push latency by nearly 5x in production and now delivers nearly a petabyte of Git data per day (OpenAI).
Notice what did not happen in either story: nobody shipped more agents to fix a problem that agents had caused. Both companies rebuilt the layer the agents depend on.
Two answers, one shared instinct
The two labs disagreed on the big design question and agreed on everything that matters for everyone else.
OpenAI kept Git. Pando is a new backend behind the existing Git protocol and client, so engineers and agents kept their existing workflows. Anthropic left Git. Scion is a new system with a new client, built on the argument that its clients (mostly agents and jobs) are always online, so Git's offline-first design was working against it.
What they share is the instinct that matters:
- Keep the workflow familiar. Even Anthropic's full migration was structured so people kept working while it happened.
- Rebuild the backend, not the habit. Both teams changed the machinery underneath the way people work, not the way people work.
- Measure everything. Both posts are dense with numbers because neither team would have known where the constraint sat without instrumenting it first.
The part that matters for your business
You are probably not merging 30,000 changes a day. But if you run agents on real work, the same physics applies at your scale.
When agents remove the bottleneck of human hands, the next bottleneck is whatever shared system all of your agents pass through: the review queue, the approval gate, the shared mailbox, the CMS, the CRM write path, the deployment pipeline. That system was sized for your team. It was not sized for your team plus a fleet of agents.
Anthropic's engineers describe their own changing role in a way that translates directly: they still review every change, but they now examine the intent of a change rather than reading every line. Their most valuable intervention, they write, was often telling Claude "no." Your reviewers, approvers, and operators are heading toward the same job description, whatever your agents produce.
Anthropic also expects this class of problem to become more commonplace over the next six to 12 months, beyond the extreme scale of its own monorepo. The window to notice your own plumbing before it notices you is open now.
What to do this quarter
- Name your shared layer. Pick the one system every agent run in your business passes through: review, approval, publishing, deployment, or intake. That is where your version of the monorepo problem will surface.
- Instrument it before it saturates. Both labs knew exactly how bad things were because they had numbers. Queue depth, wait time, and failure rate at the shared step are the metrics that matter.
- Change the backend, not the habits. When the fix comes, preserve the workflow your people and agents already trust. Swap the machinery under it.
- Move review to intent. As agent output grows, line-by-line review does not scale. Review what the change is supposed to do, and keep a human with the authority to say no.
- Expect the ceiling to move. Every bottleneck you remove will be filled. Plan capacity for the demand that shows up after the fix, not the demand you see today.
The takeaway
The lesson from Scion and Pando is not "go build a version control system." It is that automation scales until it hits the part of your operation that was built for human pace, and the fix is almost never another agent. It is rebuilding the shared step underneath everyone, keeping the familiar workflow on top, and measuring it well enough to see the next ceiling coming.
Want help finding the shared layer your agents are about to outgrow, and designing the approval and audit story around it? Talk to us.
Sources
- Redesigning source control at Anthropic (Anthropic, published October 13, 2026, updated October 15, 2026)
- Scaling Git at OpenAI (OpenAI, published November 2, 2026)
- Hacker News discussion of Scaling Git at OpenAI (November 2, 2026)
