Anthropic's Custom Source Control: When AI Changes the Code-Review Bottleneck

Anthropic's merged code changes rose from around 2,500 per day in May 2026 to more than 30,000 per day as of September, alongside about 25 million daily CI checkouts. ¹ That's a real scaling problem. It's not proof that replacing Git is the right call for anyone else.

Anthropic's custom source control: what happened and why it matters

Simon Willison, writing on October 13, 2026, describes Anthropic building two custom tools in response to the increase: Antique for code review and Scion for source control. ¹ The growth curve is steep: more than a twelvefold increase in merged changes between May and September. ¹

More changes shipping faster shifts pressure downstream. Review, validation, merge coordination, and distribution all have to keep up. That's a reasonable hypothesis about where strain could show up next. It is not what Willison's account demonstrates: it reports no review-latency or reviewer-workload numbers. ¹

Did AI assistance cause the surge? Willison guesses that Fable/Mythos-class models were a key factor, while noting that public release dates do not establish when those models became available for Anthropic's internal use. ¹ Change volume alone can't separate AI's contribution from possible effects of headcount growth, project mix, or process shifts.

The infrastructure problem Scion was built to solve

The numbers here are concrete. Anthropic's main-branch history was almost 100 GB; a compressed working copy was over 25 GB. CI performed on the order of 25 million checkouts a day, before counting developer environments, agent sandboxes, or research jobs. ¹ Fetching everything naively, Willison estimates, would mean more than 600 PB of Git traffic daily. ¹

That's a distribution problem, not necessarily a review-model problem. The figures explain why Anthropic's infrastructure team needed to address serving code for around 25 million daily checkouts. They don't establish that Git's review workflow itself was broken, and Willison doesn't provide a full technical specification for Scion. He describes it as a reinvention of centralized source control along the lines of Subversion. ¹

Worth separating: the scale justifies investigating a new system. It doesn't tell you which design was necessary, or whether another company under similar load would need the same fix.

Did AI coding make code review the bottleneck?

Anthropic makes extensive use of adversarial review, and Willison notes a specific failure mode: engineers had to steer Claude away from overengineering fixes for hypothetical concerns. ¹ That's a governance problem as much as a tooling one. More scrutiny isn't automatically better if reviewers are spending cycles on speculative risks instead of the ones that matter.

But the account gives no numbers on review time, reviewer load, defect escape rates, rework, or the split between human and model review. ¹ There's no way to tell from this whether review got slower, got better, or became the constraint as volume climbed.

The transferable lesson is narrower than “review is the bottleneck”: measure code-generation capacity and review capacity as two separate numbers. A rising merge count tells you something shifted. It doesn't tell you whether the choke point is review, CI, merge coordination, or repository transfer.

Git, Scion, or a better process: how to compare them

Three paths are worth examining. Only one is something Anthropic reports having done; the other two are what a team evaluating this should look at on its own.

Path	What to check	What the account gives you
Keep Git, fix the process	Review queues, ownership rules, work partitioning	No process-change comparison is reported 1
Keep Git, optimize infrastructure	Checkout time, repository transfer, CI throughput and cost	The scale of the problem (about 25 million checkouts per day and over 600 PB of hypothetical traffic under naive fetching), but no results from optimizing Git itself 1
Build custom source control	Supported operations, compatibility, migration risk, rollback, measured outcomes	Anthropic built the system with a seven-person team; migration took five months 1

What's missing from the public account: which Git operations Scion supports, how it handles existing tooling, and what developers gave up or gained in the switch. ¹ The seven-person team and five-month migration are signals about effort, not a cost estimate others can reuse. ¹

Before picking a path, instrument the current one. Checkout duration, merge-queue time, review age, CI cost, failed validations, and rework are measurements a team should collect for itself. Anthropic's account supplies none of them. ¹

Engineering productivity with AI: what the account doesn't measure

The daily-change figures measure merge volume. They say nothing about engineering productivity, code quality, review effectiveness, or business value. ¹ The checkout and hypothetical-traffic numbers describe infrastructure load; they don't show time saved, money saved, or whether developer experience improved under Scion. ¹ For a related look at why throughput counts are not productivity proof, see Spotify's AI coding boom as a throughput number.

Also absent are post-migration review times, throughput, defect rates, rollback incidents, adoption friction, or compatibility breakage. ¹ The AI-attribution question stays open: the account does not resolve how much of the volume increase came from model assistance versus hiring, project mix, or process change. ¹

Read plainly, this is a scale-and-architecture case study, not a controlled Git-versus-Scion benchmark. ¹ Without outcome data and a real comparison, the migration doesn't prove Scion beats Git for Anthropic, let alone for anyone else.

Migration and governance lessons for AI-generated code

A source-control swap touches more than the repository. Before committing, a team needs answers on what happens to existing tooling, integrations, developer workflows, and disaster recovery. Willison's account doesn't answer these questions for Scion. ¹

The adversarial-review detail carries its own lesson: scrutiny needs direction, not just volume. Anthropic's engineers had to steer Claude away from fixing hypothetical concerns. ¹ A review process that can't distinguish risks worth examining from risks that don't apply will burn capacity regardless of how fast code ships. Teams letting agents act on repositories may also want to read about runtime governance controls for always-on AI agents.

A practical evaluation, for anyone considering this:

• Name the actual failure mode. Slow checkouts? CI contention? Stale reviews? Merge coordination? These require different fixes.
• Baseline first. Capture the relevant workflow and quality metrics before touching any tooling.
• Test the causal claim. Confirm AI assistance is actually driving the volume increase before architecting around it; Willison leaves this open for Anthropic. ¹
• Pilot with a rollback plan. Compare against baseline; know how to revert.
• Demand operational specifics from any custom system. Supported operations, compatibility limits, failure modes, and measured effects on review time and defects.

The account does not report Anthropic following these steps; they are practical responses to gaps in the public account. ¹

What to ask before replacing Git

Separate the symptoms first. If repository transfer is the actual constraint, a new review policy won't touch it. If reviews are aging while checkouts stay fast, a source-control rewrite is solving the wrong problem. Anthropic's numbers illustrate why this distinction matters; they don't point to one universal bottleneck. ¹

Then demand evidence for the remedy, not just the problem. For any custom system, ask what it supports, how it integrates with existing tools, what can break during migration, how rollback works, and what happens to review time and defect rates afterward. None of that is in the public account of Scion. ¹

Finally, weigh the build effort against targeted fixes to the current pipeline. A seven-person team and a five-month migration provide limited information about Anthropic's staffing and elapsed migration time, not a business case for anyone else. ¹ If AI-assisted volume is also pushing up your model spend, treating AI bills as an engineering cost-control problem belongs in the same measurement effort.

The more defensible reading of Anthropic's move: it's a prompt to instrument your development pipeline when AI-assisted volume climbs, not a signal to go build a Subversion clone. Until a team can name its real constraint and show that a fix improves it without migration costs that outweigh the gain, custom source control remains an open bet, not a blueprint.

FAQ

Why did Anthropic build custom source control?

The available account says Anthropic built Scion after a sharp increase in code changes and at a scale involving an almost 100 GB main-branch history, more than 25 million daily CI checkouts, and more than 600 PB of hypothetical daily Git traffic from naive fetching. ¹

What are Anthropic's custom source-control and code-review systems called?

The code-review system is called Antique, and the source-control system is called Scion. ¹

Is Scion proven to be better than Git?

The available account does not provide a controlled Git-versus-Scion benchmark or results for review time, defects, throughput, cost, or developer experience. ¹

How long did Anthropic take to migrate from Git?

The account says a seven-person team built the new system and that migration from Git took five months. ¹

Should other companies replace Git because of AI-assisted coding?

No general recommendation follows from this account. It supports measuring repository, CI, review, and merge bottlenecks first; Scion's technical details, migration tradeoffs, and measured outcomes are not fully described. ¹

Source

¹ Simon Willison, “Redesigning source control at Anthropic,” Simon Willison's Weblog, October 13, 2026. https://simonwillison.net/2026/Oct/13/redesigning-source-control-at-anthropic/