Anthropic replaces git with Scion as agents push merges past 30,000 a day

The newsdesk feed check of 16 October 2026 surfaced 30 new videos across 58 YouTube feeds, and the sharpest item came from the Chilikanal daily AI podcast: Anthropic has rebuilt the version control behind its own monorepo. The system is called Scion, and according to an engineering blog post published on 13 October 2026 it went fully into production within five months, with the entire monorepo moved off git during that window. The trigger is a number. In May 2026 the company merged about 2,500 code changes per day, by September it was merging more than 30,000 per day, better than a tenfold increase inside four months. What makes the post worth attention is the conclusion Anthropic draws from that curve: the bottleneck in its own development is no longer the model, but developer infrastructure designed around human-scale assumptions.

The background section reads like a report from the edge of what one system can take. The history of the main branch alone runs to nearly 100 GB, a compressed working copy exceeds 25 GB, and continuous integration performs roughly 25 million code checkouts per day. Pushed through git naively, that would amount to more than 600 petabytes of git traffic per day, by Anthropic's own calculation. The defining strength of git, where every clone is a complete, self-contained repository able to compute diffs, history and blame even while offline, turns into a liability in agent operations, because almost all clients there are continuously online agents and CI jobs. The company reports it worked through the standard relief valves: shallow clones, sparse checkouts, repository snapshots kept in blob storage, and internal git mirrors, several hundred of them at peak. That bought time, but operating the setup grew harder as the merge rate kept climbing.

The post lays out six principles behind Scion that compress into two sentences: the client is always online, and the server does the work. Git treats every client as a potentially full server; Scion moves commits, merges, rebases, blame and file history onto the server. The client keeps only a bounded local object cache plus the files of its working copy, computes very little on its own, and in exchange carries far less local data weight. Every source control query also exists as an API call, and Anthropic notes that its internal pipelines for automated bisect and revert would hardly be practical at scale without that programmatic access. Scaling is horizontal at every layer: no fully materialized copy of the repository exists anywhere, including on the server, and there is no equivalent of git's packfiles, hence no repository maintenance regime. Commits are created server-side, incrementally and without locking, so their cost scales with the size of a change rather than the size of the repository. Branches and tags are deliberately kept scarce outside main and release branches, because global consensus over state becomes a scaling hazard; only the main reference moves, through a compare-and-swap on a single Spanner row.

The data structure underneath is called a changeset: an entry in a write-ahead log that records exactly the files changed relative to its first parent, plus content-addressed pointers into compressed blob storage. No packfiles, no locking, no refs to move at creation time, and that, the post argues, is what lets the system accept changes at very high throughput. Everything that is not a changeset counts as derived data: blame, path history, and skip lists for the commit graph. These structures are computed recursively from parent changesets, either eagerly or on demand, and they stay invisible to the client, which Anthropic has used repeatedly to rework internals without touching clients.

Measured against its predecessor, the new operation is drastically smaller by the company's own account: Scion infrastructure now needs about 200x less CPU and memory than the previous git mirroring setup, checkouts complete in seconds rather than minutes, which Anthropic frames as a two-order-of-magnitude relief for CI and research jobs, and continuous integration alone reported a 25 percent CPU capacity win. The architecture is a layering with one identical interface: the scn client talks over a socket to a scion-daemon on each compute node, the daemon works against regional scion-cache servers, and those stream objects straight out of RAM in the on-disk format. When something is missing there, the authoritative scion-servers answer from Google Spanner or cloud storage. All three tiers respond to the same RPCs, each reports a cache hit rate above 99 percent, and the tiering is strictly enforced to isolate failure domains, so one failure cannot take the whole system down.

The second half of the post is about building it. A team of seven people built Scion and, in parallel, a code review tool called Antique, completed the migration within five months, and did so while daily merge volume grew tenfold. One detail is already making the rounds: the first version of the pipeline that ingested git changes into Scion was handwritten. An early internal version of Claude Mythos Preview was then handed the job of rewriting the whole thing with no further instructions, and returned a version that, according to Anthropic, consumed 20x less memory, ran 4x faster, and preserved every implicit invariant of the data structures. The role of the engineers shifted along the way: increasingly their job was to tell Claude no, when agents broke core invariants or over-engineered in response to hypothetical concerns. Review now judges the intent of a change rather than every line of code. For operations, the team even deployed Claude instances as component watchers, internally called superfans, who continuously monitor telemetry and logs for anything odd.

Anyone hoping to try Scion out of the box cannot: the post states plainly that Scion is by no means a suitable replacement for git for general use. The concepts make sense, by Anthropic's own account, only for genuinely enormous monorepos, staffed by tens of thousands of engineers or by swarms of agents producing code in a stream. There is no public offering, because the surrounding ecosystem is tightly fitted to internal operation: its own code review, its own code search, its own CI. Instead, the post addresses the git community: server-side RPC architectures, ideally standardized across platforms, a negotiation protocol that no longer reconciles the full object inventory, and better composition of sparse and shallow features as starting points. Anthropic names Google's Piper plus Meta's Sapling and its Mononoke server as prior art, whose Bonsai data model pioneered several of the techniques Scion uses.

Three observation points follow for the coming months. First, whether git platforms add server-side RPC access in the spirit of Scion, which is what Anthropic proposes to the git community while explicitly treating two decades of scaling work as a foundation, not a target of replacement. Second, the timeframe Anthropic itself gives: it expects this class of problems to spread within six to twelve months, and calls it an inefficient use of engineering resources if every company on git has to solve them alone. Third, the meta-signal that after the model and the agents, developer infrastructure now sits in the bottleneck queue, from source control through builds to CI, all stated in the post. The fair counterquestion stands: every number comes from Anthropic itself, no independent comparison exists, and the company's workloads are unusual, since almost no organization writes as much code with Claude as the vendor of Claude. That transparency is still why the post matters: it is the most concrete public account so far of a daily code routine in which agents write the majority of commits.

About the author: I am Marcel, a graphic designer and operator of the HUNTER cyberdeck on a Google Pixel 6a. The newsdesk check of 16 October 2026 logged 30 new videos across 58 YouTube feeds, and the Anthropic announcement on Scion was selected as the most relevant AI story. The Chilikanal episode served purely as the entry signal, every claim in this piece traces to the linked primary source, verified on 16 October 2026.

Sources: Anthropic, Redesigning source control at Anthropic, anthropic.com, published 13 October 2026, updated 15 October 2026, accessed 16 October 2026. Simon Willison, Redesigning source control at Anthropic, simonwillison.net, 13 October 2026, accessed 16 October 2026, as secondary source to the same post. Chilikanal, episode of 16 October 2026: Anthropic Scion, Figma Specs, Agenten-Evals, YouTube, arrived in the feed check on 16 October 2026, as entry signal.