Anthropic builds custom Git replacement as AI coding velocity surges 12x

Full analysis

Anthropic replaced Git. The company built its own source control system, Scion, and a code review tool, Antique, because its internal coding bots were drowning standard tooling. The number everyone will quote is the jump from 2,500 merged code changes a day in May 2026 to over 30,000 a day by September, a 12x move in six months. The number that tells you what kind of problem this is: 25 million code checkouts per day for automated testing, against a main branch history near 100 GB.

What this means for anyone building with coding agents: the bottleneck moves. The model isn't the constraint. The GPU isn't the constraint. The plumbing around the model is. Here is where the lenses land.

The Skeptic. Thirty thousand merges a day is a great headline and a lousy measurement. Anthropic doesn't disaggregate the merges. How many are formatting passes, test scaffolds, and dependency bumps versus net-new logic that a human would have sweated over? Agents love cheap, safe, high-count changes. A 12x on merge count could be a 2x on actual shipped value, and nobody outside Anthropic can check. The adversarial review finding quietly confirms the asterisk: one model critiques another, it works, but it drifts into over-engineering and needs a human to rein it in. That is not autonomy. That is a very fast intern who gold-plates everything, and you still pay a senior engineer to say "stop."

The Safety Lens. An AI lab is now merging 30,000 machine-written changes a day into the codebase that trains, evaluates, and deploys its own frontier models. Read that twice. The human review layer is thin by the company's own account, and the pressure is to make it thinner, because thinning it is what produced the 12x. At this rate the chance of a subtle, hard-to-spot model-introduced error landing in a training or eval pipeline grows faster than any review process scales. "AI writing the infrastructure that makes AI" is happening at Anthropic today, in September 2026.

The Compute Pragmatist. Forget inference cost for a second. 25 million checkouts a day each kick off a build-and-test run. That is a private hyperscale compute loop whose workload generator is Claude, and it scales up every time the model gets better at writing code. Standard Git at that checkout rate would bleed storage I/O and network bandwidth, which is why Scion exists. The uncomfortable part for everyone else: as your agents improve, your testing bill rises faster than your capability gain. You upgrade the model, and the tax on every commit goes up with it. Most orgs have CI budgets anchored to human commit rates. Those budgets are about to look naive.

The Builder. Seven engineers, five months, Git fully migrated at this scale. That is the clearest evidence of how hard the wall is once your heaviest committer is a bot running thousands of parallel branches. Git's merge logic and lock contention were built for humans typing. They were not built for non-human commit rates, and they fall over. If you are running agentic coding loops at any serious volume, you hit this inside twelve months. You will not rip out Git. But you will start fighting merge queues, checkout storms, and a repo that balloons past what your laptop can clone. Instrument commit velocity now and know your threshold before the migration becomes an emergency instead of a project.

The Enterprise Buyer. None of this ships to you. Scion and Antique are internal, and there is no sign Anthropic plans to sell them. So the practical question for a CTO is narrower: does this change what I buy? Mostly it tells you the coding-agent story Anthropic sells is real enough that it broke their own tooling, which is a decent signal the capability isn't vaporware. But it also warns you that adopting the agent means adopting the mess downstream. Your GitHub bill, your CI vendor, your merge-queue tooling, your reviewers. Budget for the plumbing as well as the seat license.

Where they split. The Builder and Compute Pragmatist treat this as an engineering wall every serious shop will hit. The Skeptic says the wall is partly self-inflicted by counting cheap commits as if they were real work, and that most companies are nowhere near 30,000 merges a day because most companies aren't an AI lab dogfooding their own frontier model. Both are right, and that's the useful tension. The absolute numbers are Anthropic-specific and probably won't arrive at your shop for years. The direction is not. Commit velocity rises with model capability, testing compute rises faster, and the review layer gets squeezed. That pattern travels even if the 100 GB repo doesn't.

The thing that actually decides whether this matters to you: is the merge count a real productivity signal or a vanity metric inflated by trivial changes? Anthropic won't disaggregate it, because the clean 12x is the marketing and the breakdown would be smaller and messier. That silence is the answer. A lab selling coding agents has every reason to publish the number that flatters the product and no reason to publish the one that qualifies it.

Prediction: Neither Anthropic nor any other frontier lab (OpenAI, Google, Meta) will publish a breakdown of AI-generated merges by change type (trivial versus net-new logic) alongside a raw velocity figure before the next Claude major model release.

Confidence: Medium. The raw number sells the agent. The breakdown would shrink it.

Why: Anthropic put out a 12x merge-count jump with no disaggregation, in a post that doubles as a coding-agent capability pitch, while admitting model reviewers over-engineer and still need human steering. That combination says the aggregate flatters the product and the detail would qualify it. Labs publish the metric that supports the sale and skip the one that complicates it, which is why every coding-agent claim arrives as a volume or speed number and never as a "here's how much was boilerplate" number. The opposite outcome would require a lab to volunteer the less impressive cut of its own flagship story during a selling cycle, which is the behavior these posts are built to avoid.

Revisit by 2027-04-15: We're right if the next Claude major release and its surrounding materials cite coding velocity or merge counts without a published trivial-versus-substantive breakdown. We're wrong if Anthropic (or another frontier lab) publishes a disaggregated merge-quality breakdown next to a raw velocity number.