TL;DR: On October 13, 2026, Anthropic said it had moved its main code repository off Git onto Scion, a source control system it built because AI agents were merging code too fast for Git. Daily merges rose from about 2,500 in May to more than 30,000 in September, and Claude writes over 80% of the code Anthropic merges. The tool is the visible part of a larger operating model. Engineers run many Claude sessions in parallel and spend their time setting constraints and reviewing intent. PMs prototype and write evals instead of long specs. Business teams build their own tools, and Claude answers 95% of internal analytics questions. Anthropic measures adoption, throughput, autonomy, quality and accuracy, and is open about which numbers are self-reported. This article turns those practices into a blueprint other organizations can use, and is clear about the parts that do not transfer.

Most companies ask how to get employees to use AI. Anthropic is a year or two further along. It is dealing with what happens after adoption succeeds: review queues, CI capacity, source control, mentorship and a backlog of ideas larger than the company can staff. That makes it a useful preview for every organization planning AI at scale, as long as you remember that Anthropic is also the company selling the model.

What did Anthropic announce about Git and Scion?

Anthropic published “Redesigning source control at Anthropic” on its engineering blog on October 13, 2026. It describes Scion, a centralized version control system that replaced Git for the company’s monorepo, plus Antique, a new code review tool, and several related systems.

The reasons were volume and the type of client:

• Merged changes per day rose from about 2,500 in May 2026 to over 30,000 in September, roughly 12 times in four months.
• The main branch history is close to 100 GB and a compressed working copy is over 25 GB.
• CI alone performs about 25 million checkouts a day. Fetched naively, that would be over 600 PB of Git traffic daily. At peak Anthropic ran several hundred internal Git mirrors to cope.
• Almost every client is an always-online agent or CI job. Git’s main design strength, working offline with a full local copy, had become a cost.

Scion does most work on the server. Commits are “changesets” that record only what changed, queries go through an API so many CI jobs never check out code, and a three-tier cache (node daemon, regional cache, authoritative servers on Spanner) reports hit rates above 99%. Anthropic says the infrastructure footprint is about 200 times smaller than its Git mirror fleet, checkouts take seconds instead of minutes, and CI regained 25% of its CPU capacity.

The organizational details matter more for most readers than the architecture. A team of seven engineers built Scion, Antique and half a dozen related systems and completed the migration in five months, with Claude writing most of the code. An early snapshot of Claude Mythos Preview rewrote a handwritten ingestion pipeline to use 20 times less memory and run four times faster. The engineers’ job, in Anthropic’s words, was often “to tell Claude ‘no.'”

Anthropic says Scion “is by no means a suitable replacement for git for general use,” will not be open-sourced, and suits only very large monorepos. It also says other organizations should expect similar limits within six to 12 months, and suggests GitHub and GitLab adopt some of its ideas.

How much of Anthropic’s code does Claude write?

Claude authored more than 80% of the code Anthropic merged as of May 2026, according to the Anthropic Institute’s report “When AI builds itself.” That is a conservative attribution. Leadership estimates above 90% when scripts and experimental code are included. Before Claude Code’s research preview in February 2025, Claude’s share was in the low single digits.

Individual figures go further. Boris Cherny, head of Claude Code, posted on December 27, 2025 that he had landed 259 pull requests in 30 days, “every single line” written by Claude Code with Opus 4.5. At Fortune’s Brainstorm Tech in June 2026 he said he had not written code by hand in about eight months, that Claude Code is “100% written by Claude Code,” and that he manages hundreds of agents on a normal morning. In a June 2026 talk summarized by Dealroom, he estimated that loops, meaning agents prompting agents, write about 30% of his code on an average day.

Two other numbers show reliability catching up with volume. Success on open-ended tasks in Claude Code went from about 26% in late 2025 to about 91% by September 2026. And Anthropic’s staff, who rated Claude’s code below human code in late 2025, rated it roughly equal by May 2026.

How is Anthropic organized to work with AI?

Anthropic has not published an org chart, but its publications describe an organization built around small teams, shared context and agents doing most execution. Third-party estimates put headcount at a few thousand people in early 2026, which is small for a company Reuters reported had passed a $65 billion annualized revenue run rate.

Five organizational patterns recur across Anthropic’s posts and interviews:

1. Small teams own large systems. Seven engineers replaced the company’s source control in five months. Cherny’s guidance to teams is “find the bottleneck, solve the bottleneck,” then turn the fix into a reusable skill.
2. Everyone builds. Cherny has said that every member of his team writes code, including the PM, designer, finance lead and data scientist. A lawyer built a legal review plugin in an afternoon. A marketer without coding experience built a Figma plugin.
3. Agents get roles. Claude instances act as code reviewers with distinct personas and as “superfans”: dedicated monitors for single components such as the Scion cache or the merge queue, each reporting in its own Slack channel.
4. Humans own intent. Engineers set design, trade-offs and scale targets. Reviewers check the purpose of a change, and Antique’s approval rules route only consequential edits to humans.
5. Culture is managed on purpose. New engineers report talking to Claude more than to colleagues, so Anthropic encourages pair programming and social time. Cherny says engineers need to “feel very safe being wrong.”

How do Anthropic engineers work with Claude Code?

Anthropic engineers work as orchestrators of several Claude sessions at once. The workflow below comes from Boris Cherny’s public posts on X during 2026 and Anthropic’s own documentation.

Run work in parallel. Cherny ran five Claude sessions in numbered terminal tabs plus five to ten on the web in January 2026. The team’s top tip is three to five git worktrees, each with its own session. By mid-2026 “dynamic workflows” could fan a large migration out to hundreds of subagents.

Keep one shared memory file. Each team keeps a CLAUDE.md file in the repository. “Anytime we see Claude do something incorrectly we add it to the CLAUDE.md, so Claude knows not to do it next time.” Engineers tag @claude in pull request reviews so rule updates ship with the code.

Turn repetition into skills. Anything done more than once a day becomes a slash command or skill checked into git, such as /commit-push-pr, /simplify or /go, which tests end to end, simplifies and opens a PR.

Make Claude verify its work. Cherny says giving Claude a way to check its output (tests, a browser, a phone simulator) improves quality two to three times. In September 2026 he added that “production code written by Claude should have a higher bar than if it was written by a human,” backed by lint rules, end-to-end tests and daily Claude-powered fuzzers.

Automate the routine. Loops and scheduled routines babysit pull requests, prune stale branches and summarize Slack. /goal keeps Claude working until a stated condition is met.

Expect practices to expire. In January Cherny started most sessions in plan mode. By June he had largely dropped it because newer models plan well on their own. Workflows are reviewed as models change.

How do product managers work at Anthropic?

Anthropic PMs prototype and write evals. Scott White, product lead for Claude.ai, describes his workflow as writing the PRD, generating evals from it and implementing those evals in the product. His summary: “evals are the new PRD.” Cat Wu, product lead for Claude Code, says evals “make an abstract product feel more concrete,” and uses Claude Code for prototypes and evals, Claude.ai for strategy and judgment calls, and agentic tools for the rest.

Three practices stand out:

• Technical PMs. Wu has said almost all PMs on her team have engineering backgrounds or have written code with Claude Code.
• Side quests. Several Claude Code features, including the desktop app, the AskUserQuestion tool and the to-do list, came from self-directed experiments rather than the roadmap.
• More product capacity. Cherny argues that once code is mostly AI-written, the bottleneck moves upstream to deciding what to build and downstream to go-to-market. His answer is more PMs and user researchers.

How do non-engineering teams at Anthropic use AI?

The most documented example is analytics. In June 2026 Anthropic’s data team reported that Claude handles about 95% of internal business analytics queries with about 95% accuracy overall, approaching 99% in some domains. Without encoded skills and business context, Claude answered only 21% correctly. The gain came from four layers: governed data models and metrics, a knowledge layer with semantic definitions and lineage, skills that encode repeat analyses, and validation checks. The data scientists moved to causal modeling, forecasting and machine learning.

Other teams follow the same pattern of encoding expertise once and reusing it:

• Legal. Associate general counsel Mark Pike built a review plugin from markdown files and the team’s playbooks. Users paste a draft, Claude flags issues by risk level and posts a summary to the legal team in Slack. Lawyers still review all output. Marketing review turnaround fell from two to three days to 24 hours, according to Anthropic’s agentic coding trends report.
• Growth marketing. A workflow takes a CSV of hundreds of existing ads and generates new variations in minutes. A Figma plugin creates up to 100 ad variations per batch.
• Finance and product. Mike Krieger describes packaging one finance employee’s contract review method as a skill so colleagues get the same quality. PMs query data directly through Claude instead of waiting for dashboards.
• Security and infrastructure. Security engineers trace incidents from stack traces about three times faster. In April 2026 Claude shipped over 800 fixes that cut a class of API errors by a factor of 1,000, work the overseeing engineer estimated would have taken a human about four years.

What KPIs does Anthropic measure?

Anthropic tracks five kinds of metrics: adoption, throughput, autonomy, quality and accuracy. It labels survey-based numbers as self-reported and warns they may overstate gains.

KPI category	What Anthropic measures	Example values
Adoption	Share of work done with Claude; teams and task types	28% → 59% (2024–2025)
Throughput	Merged PRs per engineer per day; merges per day; lines merged	+67% PRs; 2,500 → 30,000 merges/day
Autonomy	Consecutive tool calls without a human; human turns per session	9.8 → 21.2 tool calls; 6.2 → 4.1 turns
Quality	Task success rate; bugs caught by Claude review; staff ratings of code quality	26% → 91% success; review would catch ~1/3 of past incident bugs
Accuracy	Correct answers on business questions with and without skills	21% → 95%+
New value	Share of AI work that would not otherwise exist	27%
People	Delegation share, skills concerns, mentorship and collaboration	Most can fully delegate only 0–20% of work

Anthropic is candid about the weak spots. Its December 2025 study notes that productivity was self-reported and may reflect social-desirability bias. The Anthropic Institute says lines of code overstate gains and that the true research uplift was “probably somewhat lower” than staff’s median estimate of about 4x. Its own introduction and body text even give inconsistent versions of the “8x code per engineer” figure. Copy the metric categories, and copy the habit of labeling evidence quality too.

What results and side effects has Anthropic seen?

The benefits are well documented: more output per engineer, more small fixes and tools that would never have been prioritized, faster incident response, and business teams that answer their own data questions. One incident fix that normally takes two to three days took about two hours. Merged pull requests per engineer rose 67% after Claude Code rolled out across engineering.

The side effects are documented too, and they matter for any blueprint:

• Review becomes the bottleneck. Anthropic says human code review is now a constraint, which is why it built multi-agent Claude review and intent-based approval rules.
• Volume breaks infrastructure. Source control, CI and merge queues were designed for human-speed work. Scion is the result.
• Skills and mentorship erode. Some engineers worry about losing deep coding skills and practice without AI on purpose. One employee said 80–90% of their questions now go to Claude, and juniors report fewer chances to learn from colleagues.
• Ideas outrun capacity. Anthropic reports more initiatives than it can pursue. Choosing and stopping work becomes the scarce skill.
• Work changes shape. Respondents describe shifting from writing code to reviewing it, with one estimating that 70% of their work is now review and revision.

A blueprint for implementing AI at scale

The following eight steps translate Anthropic’s practices into a sequence a normal organization can follow. Each step has a KPI so progress is visible.

Step	What Anthropic does	What you can do in 90 days	KPI to track
1. Shared context	CLAUDE.md per repo, updated after every mistake	One context file per repo and team, owned and reviewed like code	% of active repos with a maintained context file
2. Skills	Anything done daily becomes a checked-in skill	Pick the 10 most repeated workflows per function and encode them	Skill usage per week; time per task
3. Verification first	Claude tests its own work; fuzzers; E2E tests	Require automated tests or checks before agent changes merge	% of agent changes with automated verification; escaped defects
4. Intent review	Humans review purpose; risk-based routing	Tier changes by risk so humans review the ones that matter	Review time; % of changes needing human review
5. Governed data	Semantic layer, skills, validation for analytics	Define canonical metrics before letting agents query data	Accuracy on a fixed evaluation set of business questions
6. Builders everywhere	Legal, marketing, finance build their own tools	Give business teams agent access with clear accountability rules	Tools built by non-engineers; turnaround times
7. Bottleneck reviews	“Find the bottleneck, solve the bottleneck”	Monthly review of where work queues (review, CI, approvals, decisions)	Cycle time by stage; queue length
8. People safeguards	Pair programming, social time, practice without AI	Protect mentoring time for juniors; track skills and wellbeing	Mentoring hours; engagement; skills assessments

What can’t other companies copy from Anthropic?

Several conditions at Anthropic are unusual, and they limit how directly its numbers transfer.

• Model access. Anthropic staff use internal models before release, including early Claude Mythos Preview snapshots. Outside companies work with released models and rate limits.
• Cost tolerance. Cherny reported using over 8 billion tokens in a few months. The Register notes that rebuilding source control is easier “when cost is no object.”
• Self-selected staff. Anthropic hires people who want to work this way. Most organizations have mixed skills and mixed enthusiasm.
• Product overlap. Anthropic sells Claude, so every internal success is also marketing. Its own reports are careful, but the incentive exists.

What does transfer is the method: shared context, skills, verification before volume, human review of intent, governed data, and a regular habit of finding the next bottleneck.

FAQ

How does Anthropic use AI internally? Anthropic uses Claude across engineering, research, product, data and business teams. Claude writes over 80% of the code Anthropic merges, answers about 95% of internal analytics queries, reviews code, monitors systems and powers tools that legal, marketing and finance teams built themselves. Humans set goals, constraints and review intent.

Why did Anthropic replace Git? AI agents increased merges from about 2,500 to over 30,000 per day between May and September 2026, and CI performed about 25 million checkouts daily. Git’s design, built around full local copies, could not serve that load efficiently. Anthropic built Scion, a centralized, server-side system that cut infrastructure use by about 200 times.

What is Anthropic Scion? Scion is Anthropic’s internal source control system for its monorepo, announced October 13, 2026. It stores commits as changesets, runs queries on the server, supports sparse checkouts and uses regional caches. Anthropic says it is not a general Git replacement and will not open-source it.

How much of Anthropic’s code is written by Claude? More than 80% of merged code as of May 2026 by Anthropic’s conservative measure, and over 90% including scripts and experiments by leadership estimates. Boris Cherny says Claude Code itself is 100% written by Claude Code.

How does Boris Cherny use Claude Code? He runs many sessions in parallel, keeps a shared CLAUDE.md file updated after every mistake, turns repeated tasks into skills, has Claude verify its own work, and uses loops and routines for background tasks. By mid-2026 he reported managing hundreds of agents and no longer writing code by hand.

What KPIs does Anthropic use to measure AI impact? Share of work done with Claude, merged pull requests per engineer, merges per day, autonomous tool calls and human turns per session, task success rate, bugs caught by AI review, analytics accuracy and the share of AI-assisted work that would not otherwise exist. It labels survey-based figures as self-reported.

How productive are Anthropic engineers with AI? Merged PRs per engineer per day rose 67% after Claude Code adoption. Self-reported productivity gains rose from 20% to 50% in a year. Research staff estimated a median 4x output increase in March 2026, though Anthropic says the true figure is probably lower.

How do Anthropic product managers use AI? They prototype in Claude Code, generate evals from product requirements (“evals are the new PRD”), query data directly and run side experiments that sometimes become features. Most Claude Code PMs have engineering backgrounds.

Can other companies copy Anthropic’s approach? They can copy the method: shared context files, skills, automated verification, risk-based review, governed data and bottleneck reviews. They should not expect Anthropic’s numbers, which depend on early model access, high spending and a workforce selected for AI-native work.

What problems has AI created at Anthropic? Code review and source control became bottlenecks, ideas outgrew capacity, and some employees worry about skill loss and fewer mentoring opportunities for juniors. Anthropic responded with AI review, Scion, pair programming and social time.

Final take

Anthropic’s Git replacement is a symptom. The cause is an operating model where agents do most of the execution and people spend their time on context, constraints, verification and judgment. That model produced a twelvefold jump in merges in four months, and it moved the company’s problems from “how do we get people to use AI” to “what do we build next, and who reviews it.”

Most organizations are still on the first rung of that ladder. Anthropic’s useful lesson for them is the order of the work: build shared context and verification before you scale generation, measure bottlenecks instead of activity, and protect the people who are learning a new job.

Sources: Anthropic Engineering, “Redesigning source control at Anthropic” (October 13, 2026); Anthropic Institute, “When AI builds itself” (2026, updated September 18); Anthropic, “How AI is transforming work at Anthropic” (December 2, 2025); Claude blog, “How Anthropic teams use Claude Code” (July 2025); Anthropic, “2026 Agentic Coding Trends Report“; InfoQ on Anthropic’s self-service analytics (June 2026); DevOps.com and The Register on Scion; Fortune on Boris Cherny (June 11, 2026); Simon Willison’s Boris Cherny archive; paddo.dev summary of Cherny’s workflow thread; Dealroom summary of Cherny at Scale; Fast Company; The Product Folks on Anthropic PMs. Most figures are published by Anthropic about itself. Survey-based productivity numbers are self-reported. Some social media figures come from secondary summaries and were not checked against original recordings.