The Signal / AI in practice / October 19, 2026
Anthropic replaces git with Scion
Anthropic redesigned source control for agents appraisal review faces the same choice
Anthropic reports that daily merges to its main monorepo grew from about 2,500 in May 2026 to over 30,000 by September, a 12x increase in four months. Git’s offline, distributed design stopped fitting, so a seven-person team spent five months building a server-first replacement called Scion. The pattern, redesigning a process instead of automating it, is the same choice appraisal review faces as AI drafting volume rises.
Anthropic reports
- Daily merges, Sept. 2026
- 30,000
- Built by
- 7 people
- In
- 5 months
- Smaller footprint
- 200x
Merge volume outran git in four months
Anthropic reports daily merges grew 12x, from about 2,500 to over 30,000
Anthropic’s engineering blog disclosed Scion, an in-house version control system that replaced git for the company’s main monorepo, on October 13, 2026, with an update on October 15 crediting prior art it had first left out. Anthropic reports the team was seven engineers, and the build took five months, against roughly a tenfold increase in merge volume over the same span. A companion code review surface called Antique replaced pull requests.
Much of that merge volume is Claude’s own commits. The engineers who built Scion rate Claude better at writing code than themselves. The rest of this article turns on what that rating leaves out: the job also includes reviewing, coordinating and deciding what to build next, and Anthropic’s engineers still do all three.
AI-generated illustration: merge volume accelerating, the way index cards accelerate into a catalog.
Source: Redesigning source control at Anthropic ↗, Anthropic Engineering, Oct. 13, 2026.
A design built for offline humans, not online agents
Git assumes every clone is a complete, offline copy of the repository
Anthropic reports its main branch history runs near 100 GB, with a compressed working copy over 25 GB. Git’s model gives every client a full, independently usable copy so it can work offline, which made sense when a clone was a person’s laptop. Anthropic reports about 25 million CI checkouts a day alone, not counting development, sandbox or research checkouts.
A naive git setup moving a full working copy on every checkout would move 625 petabytes a day (25 million checkouts times 25 GB). Anthropic ran several hundred git mirrors at peak to absorb that load, and the mirrors still drifted from main between refreshes, a gap Anthropic calls seeds. The mirror fleet was itself a symptom: a distributed design pressed into always-online, agent-scale service.
What a full-clone checkout costs at scale
Anthropic’s own figures, recomputed live.
625petabytes a day, moved by a naive full-clone checkout
Source: Redesigning source control at Anthropic ↗. Checkout figure excludes development, sandbox and research checkouts.
Scion moves history, blame and merges to the server
Clients request only the trees and blobs they need; they never declare what they already have
Scion’s unit of work is a changeset: a write-ahead log entry listing the files a change touches against its first parent, with content-addressed pointers to the actual file contents. Everything else, blame, path history, search indices, is derived data computed from changesets, either eagerly or on demand. A per-machine daemon serves every job on that machine from a shared object store; a regional cache serves the daemon; an authoritative server tier reads the backing store and computes anything the cache does not already have. Anthropic reports a cache hit rate above 99% at each tier, so the rare miss that reaches the server is the exception, not the rule.
Snapshots the whole tree, every file, every commit.
Lists the 3 files this change actually touched.
AI-generated illustration: requests mostly stop at the cache; the rare miss travels on to the server.
A 200x smaller footprint, by Anthropic’s own measure
Every figure in this chapter is Anthropic’s own report; none is independently audited
Anthropic reports Scion’s footprint is 200 times smaller in CPU and memory than its git-mirroring setup, that checkouts dropped from minutes to seconds, and that the move freed 25% of CI CPU. Its one detailed worked example: an early Claude Mythos Preview snapshot rewrote the git-to-Scion ingestion pipeline itself, using 20 times less memory and running 4 times faster. Anthropic’s own words for the pressure behind all of it: the latent demand for merging code is practically infinite.
What outside readers pushed back on
Mononoke/Sapling credit
A contributor to Meta’s Mononoke said Scion’s derived-data design follows Mononoke’s Bonsai model; the original post credited only Sapling. Anthropic added a footnote on Oct. 15.
Local history editing
Server-side commits remove the private local history editing a distributed system allows. Unresolved in public discussion.
Merges per employee
Some commenters called 30,000 merges a day modest once divided across Anthropic’s headcount. Unresolved.
A single point of failure
A mandatory central service concentrates risk. Anthropic points to directory-level IAM as the mitigation.
Lines no human reads
Intent review, reviewing the purpose of a change rather than every line, drew the objection that some lines ship unread. Anthropic says humans review every change request.
Sources: Anthropic ↗ Simon Willison ↗ Hacker News ↗ Mononoke credit discussion ↗
Claude writes the code; people check what each change is for
Humans still review every change request, for its purpose rather than each line
Anthropic could not effectively replicate a global merge-queue lock at this volume, so it invested in post-merge and pre-deploy testing instead: changes merge, and Claude fixes the conflicts that turn up, rather than every change waiting on a single queue. Directory-level permissions and semantic approval rules route attention toward sensitive changes and let comment-only edits move faster. One Claude agent per component, which Anthropic calls a superfan, watches telemetry and logs for its own area.
Comment-only
- Documentation edits
- Formatting and style fixes
- Test additions with no logic change
Logic-affecting
- Changes to sensitive directories (IAM-gated)
- Anything crossing a semantic approval boundary
- Intent reviewed by a human before merge
AI-generated illustration: identical copies piling up is what review without triage looks like at this volume.
Humans review every change request’s intent; Anthropic says adversarial Claude-on-Claude review is effective but sometimes over-engineers against threats that are not real.
Redesign beats retrofit, and gains arrive late
Hammer’s 1990 argument and David’s dynamo both describe the same lag: a new technology pays off once the process around it is redesigned
In 1990, management researcher Michael Hammer argued that companies kept using computers to speed up existing paperwork rather than rethinking the process the paperwork served. His phrase: It is time to stop paving the cow paths.
Economist Robert Solow made a related point in 1987, reviewing a book on manufacturing competitiveness: You can see the computer age everywhere but in the productivity statistics. Computers were everywhere in offices by then, but measured output per worker had not moved. Paul David’s 1990 paper on factory electrification gives the historical reason why: factories first swapped the steam engine for a large electric motor and kept the old line shafts and belts, and the productivity gains stayed small. They arrived in the early 1920s, four decades after the first central power station opened, once factories gave each machine its own electric motor and rebuilt the floor around that change.
Sources: Hammer, "Reengineering Work" ↗, Harvard Business Review, July/Aug. 1990. Solow, review of Manufacturing Matters ↗, The New York Times Book Review, June 12, 1987. David, "The Dynamo and the Computer" ↗, American Economic Review vol. 80 no. 2, May 1990, pp. 355 to 361; the four-decade span is on p. 357. Brynjolfsson, Rock and Syverson, "The Productivity J-Curve" ↗, NBER working paper 25148, 2018; American Economic Journal: Macroeconomics 13(1), Jan. 2021.
The same choice, a different desk
Appraisal review, coordination and revision were built around a person at every step
Review becomes the constraint once drafting gets cheap
The engineers who built Scion rate Claude better at writing code than themselves. That moved their bottleneck from typing to reading: someone still has to go through every change line by line before it merges.
An appraisal report that an AI tool drafts faster than a human can read creates the same shift. If a reviewer still reads every adjustment, every comp swap and every condition rating at the same pace as before, the bottleneck has moved from drafting to review. It is still on the same desk.
The dynamo trap
UAD 3.6 redesigns the data a report carries. The workflows around it, who orders a report, who revises it, who reviews it, still assume a person at each handoff. Adding AI drafting to a workflow built around a person at every step is the retrofit Paul David described: one motor bolted onto the old shafts and belts.
The gains from electrification arrived once factories gave each machine its own motor and rearranged the floor around it. The equivalent for an appraisal shop is rearranging how work and review move between people. Giving each person a faster drafting tool is the easier step, and the smaller one.
Coordination is most of the turn time
Order assignment, fee and due-date negotiation, revision requests and reconsideration-of-value routing are mostly information passing between people, plus waiting. None of it is drafting.
Anthropic’s engineers replaced a global merge-queue lock, which made every change wait for a coordinator to clear it, with post-merge testing: changes go in and Claude fixes the conflicts that come up, instead of a human gatekeeping each one. An appraisal shop’s equivalent coordinator roles, assigning orders, routing revisions, chasing a due date, are candidates for the same kind of change before anyone argues about which roles should shrink.
AI-generated illustration: sorting work into two trays is the triage every one of these coordination steps needs.
Engineers still tell Claude no
What an appraiser contributes maps onto what an engineer contributes, reviewing intent instead of typing faster
Anthropic’s engineers still tell Claude no. They still review intent rather than every line, which only works because a human is accountable for the result.
What a reviewer brings to that exchange maps onto what an appraiser brings to a report: the signed certification that makes one person accountable, the scope of work that sets the assignment’s goal, the standing with a lender, borrower or AMC that trust is built on, and the unwritten knowledge of a local market that no model has absorbed. Automating the coordination around a job does not remove the need for someone to stand behind its conclusion.
Anthropic’s own forecast, narrower than the headlines
Anthropic says this problem class reaches other large monorepos in 6 to 12 months, not that every company needs a Scion
Anthropic’s own claim is narrower than some secondary coverage suggested: Scion fits monorepos at a scale most companies do not run, and the company expects the underlying problem, agent-scale, always-online source control traffic, to reach other large monorepos within 6 to 12 months. It has also said it would rather see git itself adopt compatible server-side and protocol changes than publish Scion for general use.
For an appraisal operation, the item to watch is not whether a Scion-like tool arrives. It is whether AI-drafted volume starts to outrun the review, coordination and revision steps built around one person at a time, the same way merge volume outran git at Anthropic. That pressure shows up first in backlog and turnaround time, before anyone decides to redesign the workflow around it.
AI-generated illustration: the same field as the opening shot, now with the planned road finished.
Watching drafting volume rise in your own shop?
Tell us where AI drafting has outrun your review or revision process.We are tracking which coordination steps shops are redesigning first, and which they are leaving as is.→This article follows Anthropic’s own published account and a secondary video summary; every metric is Anthropic’s self-reported figure, not an independent audit. Charts in this article are recreated from the described shape of Anthropic’s published chart and from the cited economics papers, not copies of Anthropic’s own images. Video illustrations are AI-generated and do not depict Anthropic’s offices, products or people. This article is general information, not appraisal, legal or technology advice. See how deep appraisers take AI under Advisory Opinion 41 for the appraisal-specific disclosure rules this translation assumes.
The Signal / Monthly briefing
Follow AI in appraisal practice, with sources.
One sourced briefing in the first week of each month.
Unsubscribe in one click. We don't sell your email or share it with sponsors. Privacy Policy