Agent throughput selected for server-centric source control and post-merge verification
Source → Observed → Interpretation → Model implication
Redesigning source control at Anthropic
View source →Anthropic reports that average changes merged per day increased from around 2,500 in May 2026 to more than 30,000 in September, after which source control became its largest development bottleneck.
The main-branch history of Anthropic's monorepo is reported to be almost 100 GB, while a compressed working copy exceeds 25 GB.
Anthropic performs approximately 25 million code checkouts per day for CI, before developer environments, agent sandboxes, and research jobs. Naively fetching the full compressed working copy would imply more than 600 PB of daily Git traffic.
Anthropic used shallow clones, sparse checkouts, promisor remotes, repository snapshots, and several hundred internal Git mirrors before concluding that this architecture no longer scaled with the repository and merge rate.
Anthropic built Scion and fully migrated its monorepo away from Git in five months. The project also required a new code-review surface, code search, source-control identity and access management, and new mechanisms for landing large volumes of reviewed code.
Scion assumes clients are online, moves history, blame, commit, merge, and rebase operations to servers, and gives clients bounded local object caches rather than complete repository copies.
Scion exposes source-control queries through APIs so CI jobs and internal tools can inspect diffs, ancestry, blame, and related metadata without maintaining a working copy or local repository.
The system represents commits as server-created changesets containing the files changed relative to a parent and content-addressed references to file contents. Creating changesets does not require packfiles, locks, or moving a shared reference.
Anthropic discourages references other than main and release references because global state requiring cross-server agreement becomes a scaling constraint. Updating a supported reference is separated from creating the changeset.
Scion stores additional history, blame, search, and commit-graph structures as derived data that can be computed eagerly or on demand without changing clients.
Immutable content-addressed objects allow any server or cache tier to answer requests without checking for mutations. Anthropic reports cache hit rates above 99% at each tier and independently scalable, short-lived components.
Anthropic reports that Scion uses 200 times less CPU and memory than its previous Git-mirroring infrastructure, while CI and research checkouts are two orders of magnitude faster, moving from minutes to seconds.
Anthropic reports that the faster checkout architecture freed substantial fleet capacity and produced a 25% CPU-capacity gain for CI.
A seven-person engineering team built Scion, its Antique code-review system, and several related services and completed the migration in five months while merged-change volume increased more than tenfold.
Anthropic reports that an early Claude model rewrote its Git-to-Scion ingestion pipeline to use 20 times less memory and run four times faster while preserving implicit ordering and data-structure invariants.
Engineers report continuing to review every change request, but increasingly reviewing the intent rather than every generated line. They describe their role as defining high-level system constraints and rejecting locally plausible implementations that violate those constraints or add unnecessary complexity.
Anthropic assigns persistent monitoring agents, called superfans, to individual Scion and Antique components. These agents inspect telemetry and logs for subtle anomalies alongside conventional alerts.
Anthropic reports that a conventional globally locked merge queue has not been effective at its change rate. It is investing instead in post-merge and pre-deployment automated testing while allowing concurrent changes and using Claude to repair conflicts.
This is strong production evidence for Selection. Agent-enabled implementation increased the rate of change until source control, checkout capacity, review, integration, and CI became the limiting environment. Anthropic then selected an architecture whose fitness assumptions match online agents and ephemeral jobs: server-side operations, sparse and optional working copies, immutable cached objects, incremental lock-free changesets, minimal global references, and independently scalable services. It refines Cooperation by showing that agent-scale composition depends on minimizing the state that requires global agreement rather than coordinating every action synchronously. It also refines Specialization: engineers move toward architectural constraint and intent review, while persistent monitoring agents specialize by component, but the source does not quantify the superfans' independent contribution.
REFINES. Supports C02 with a completed production migration and measured changes in resource use, checkout latency, and CI capacity after the previous substrate became a bottleneck. Refines C03 by showing that higher-throughput cooperative work selects for less synchronous global coordination and more shared server-side capabilities. Refines C04 with evidence of differentiated human architectural judgment and component-specific operational agents, while stopping short of claiming a measured higher-order fitness advantage from those specialized roles.
What this does not establish
- The article is a first-party engineering account and does not provide an independently audited dataset or reproducible performance benchmark.
- Anthropic reports total merged changes but does not quantify how many were authored by agents, how large they were, or how much customer or research value they produced.
- Merge count, checkout count, code volume, and infrastructure throughput are activity measures rather than equivalent productivity, software quality, or business outcomes.
- The 200-fold infrastructure reduction, two-orders-of-magnitude checkout improvement, and 25% CI capacity gain are point estimates without published distributions, baselines, measurement windows, uncertainty, or independent replication.
- Scion was built for an exceptional monorepo and agent workload. Anthropic explicitly states that it is not a suitable general replacement for Git and that the organization can spend more tokens internally than most businesses.
- Migrating from Git required replacement or adaptation of code review, code search, identity and access management, CI, merge, and other development- lifecycle systems; the article does not provide the full migration cost.
- The seven-person, five-month delivery timeline does not isolate Claude's causal contribution from the team's expertise, prior systems, organizational investment, or unusually favorable access to Anthropic models.
- The reported 20-fold memory and fourfold speed improvements concern one internally selected ingestion rewrite and do not establish general agent superiority in infrastructure engineering.
- Reviewing change intent rather than every line does not establish review depth, defect escape rate, security outcomes, rollback rate, or production reliability at the higher merge volume.
- Moving verification after merge changes where integration risk is absorbed; the source does not quantify conflicts, main-branch failures, recovery time, or the effectiveness of pre-deployment containment.
- The component-specific superfan agents are described without comparative detection, false-positive, incident, or reliability measurements, so they do not alone establish C04's higher-order fitness advantage.
As agent-produced change exceeds human-paced integration capacity, which source-control, review, verification, and interface-ownership designs preserve useful throughput without moving unacceptable failure risk into mainline code or production deployment?