Source Control and Code Review at Agent Scale

When agents write most of the code, the bottleneck moves to infrastructure built for humans. Two Anthropic posts from Oct 13–15 show one answer for source control and one for code review, and a community response sharpened the question of prior art.

Part of Weekend Deep Dive #1: Agent-Scale Infrastructure.

---

What’s new this week

On Oct 13 Anthropic described Scion, an in-house version control system that replaced git for its monorepo in five months. Two days later the Claude Code team explained how it reviews code when, according to the team, Claude writes about 90 percent of Claude Code’s own code. Both posts start from the same number: according to Anthropic, its monorepo went from about 2,500 merged changes a day in May 2026 to more than 30,000 a day in September.

Why it matters. Few teams are near that volume, but the direction applies to any organisation that puts coding agents in front of many engineers. Checkout cost, CI capacity and human review time become constraints long before a custom VCS would make sense. Anthropic itself expects “this class of problems will become more commonplace over the next six to 12 months.”

The week also produced a prior-art debate. A developer who says they helped build Meta’s open-source Mononoke server argued that Scion’s core data model closely follows it, and on Oct 15 Anthropic added a footnote crediting Mononoke (details below).

Key ideas

Source control (Anthropic, Oct 13)

• Always-online clients. Nearly all of Anthropic’s clients are agents and CI jobs that are online by definition, so git’s offline-first design worked against it. Scion moves history, blame, merges and rebases to the server, and the client keeps only a bounded cache.
• Source control as a service. Every query the client uses is also an API, so many jobs need no local checkout at all. Anthropic says its automated bisect and revert pipelines depend on this.
• Changesets plus derived data. A changeset records only what changed relative to its first parent. Blame, path history and similar indexes are computed from changesets by one generic mechanism.
• Tiered, content-addressed caching. A per-node daemon, regional caches and authoritative servers answer the same RPCs. Anthropic reports a cache hit rate above 99% at each tier and a footprint 200x smaller in CPU and memory than its old git mirroring setup. For scale: the post counts about 25 million checkouts a day for CI alone, before developer environments, agent sandboxes and research jobs.
• No merge queue at this rate. Anthropic writes that a proper merge queue “doesn’t make much sense at agent scale” and relies on post-merge and pre-deploy testing instead. It is explicit that Scion is “by no means a suitable replacement for git for general use.”

Code review (Claude Code team, Oct 15)

• The goal is human time, not human reading. Most PRs should merge with at most one human in the loop, who can be the author, with revert and incident rates as the guardrail.
• Risk tiers set by the change, not the author. A classifier picks low, medium or high risk. Sensitive paths always need a person, and “an agent’s approval never counts as a person’s.”
• Findings verified by other agents. Parallel bug-finders, then a voting pass. When the research preview launched, PRs with substantive comments went from 16% to 54%, according to the post.
• Review loops must converge. One PR went through more than fifty overnight rounds. About 59% of what the reviewer kept flagging was code that predated its first pass. The fix is a three-round cap and diff-only re-review.
• Incidents become evals. Past incident-causing PRs form a daily eval set for the reviewer. The team says it does not yet have circuit breakers that tighten approval when revert or incident numbers move.

flowchart LR
    V["Merge volume<br/>~2,500 → 30,000+/day<br/>(Anthropic, Oct 13)"] --> S["Source control<br/>Scion: server-side history,<br/>tiered caches"]
    V --> R["Code review<br/>risk tiers, agent-verified findings,<br/>round caps (Claude Code team, Oct 15)"]
    M["Prior art: Mononoke / Bonsai changesets<br/>(Meta, open source)"] -. "credited in Oct 15 update" .-> S
    C["Cursor Continuity<br/>S3 write-ahead log under git<br/>(background, Aug 2026)"] -. "alternative: keep git" .-> S
    R --> G["Guardrail: revert and incident rates"]

Analysis: the two posts describe one shift from opposite ends. Scion cuts the cost of each change; the review process cuts the human attention each change needs. Teams far below Anthropic’s scale can borrow the second half (risk tiers, round caps, incident evals) well before the first.

The prior-art debate

On Oct 14, Rain of sunshowers.io, who says they helped build Mononoke at Meta, argued that Scion closely mirrors it and that the missing credit “borders on plagiarism”, while stating no objection to the design itself being copied. A follow-up on Oct 15 alleges the copying was deliberate; the author says they privately received confirmation, and the claim has not been independently confirmed. Anthropic’s post now has a footnote naming Mononoke as having “pioneered many of the techniques we use in Scion, in particular the Bonsai data model.” The follow-up’s top note says “sufficient attribution has been added,” though its body still says no update had been seen when it was written.

Analysis: as Rain notes, agents that help design systems “are likely to index heavily on existing designs.” A prior-art check in design review is a cheap habit to adopt.

Reading list

• Redesigning source control at Anthropic — Anthropic Engineering, Oct 13, 2026 (updated Oct 15). The primary source on Scion: the six design principles, the changeset and derived-data model, the checkout architecture, and the lessons on engineers telling Claude “no.” The closing section suggests how git forges could adopt some of the ideas without leaving git.
• How the Claude Code team does code review — claude.dev, Oct 15, 2026. The full review pipeline with the risk-tier table, the convergence fixes and an adoption checklist (before the PR, on every PR, approval, merge and after) aimed at teams that let agents open pull requests.
• Hacker News discussion of the Scion post — Hacker News, Oct 13, 2026. Over 1,000 comments as of Oct 17, covering practitioner reactions to the scale claims, code quality, and the phrase “examine the intent of the change.”
• Anthropic’s Scion is heavily inspired by Mononoke — sunshowers, Oct 14, 2026. Community response. A short comparison of how Anthropic and the Mononoke team each describe derived data. The follow-up post (Oct 15) explains the Bonsai changeset model and why it was contrarian in 2018.
• Hacker News discussion of the Mononoke post — Hacker News, Oct 14, 2026. Counter-arguments (Mononoke is part of Sapling, which was credited) alongside the case for naming it. Includes a useful explanation of how Sapling, EdenFS and Mononoke fit together.

Background and prior art

• Mononoke — source in the Sapling repository (Meta, open source). The server whose Bonsai changeset model Scion’s changesets follow, per Anthropic’s updated footnote.
• Sapling — sapling-scm.com (Meta, open source since 2022). The client and wider system that Anthropic originally credited.
• Git at any scale — Cursor, Aug 18, 2026. A different answer to agent-driven repository load: keep git, but make an S3 write-ahead log the source of truth and treat local repositories as caches.