Anthropic’s Scion: scaling agent-produced changes without abolishing human decisions

Source: Nick G. and the source-control team, “Redesigning source control at Anthropic,” October 13, 2026; Willison’s October 13 interpretation is a pointer, not an independent audit. The work spans roughly five months of building and migration through September, not a one-day release. Anthropic says daily merged code changes rose from ~2,500 in May to >30,000 in September. It does not define those as 30,000 independently requested features or identify the counterfactual cause. The team designed Scion (version control) and Antique (code review) for an unusually large monorepo and many online agent/CI clients. About 25 million daily CI checkouts, nearly 100 GB of mainline history and a >25 GB compressed working tree made Git mirror maintenance a bottleneck. Server-side content-addressed changesets, bounded client caches, sparse and optional checkouts, tiered object caching and query APIs replace maintaining full repo copies in many consumers; the reported source-control infrastructure CPU/memory footprint is 200x smaller than its prior Git mirror setup, with a 25% CI CPU-capacity gain. These are first-party infrastructure observations, not model productivity trials.

Human boundary and a non-obvious trade-off. Seven engineers built and migrated Scion, Antique and related systems in five months. They say they review every change request for intent, not line-by-line code. Their interventions include rejecting agent code that scans a commit log instead of calling the review service and stopping adversarial reviews that added needless defensive validation. Antique can encode semantic approval requirements so humans focus on consequential logical changes; source-control access can be scoped to sensitive directories. The team has not made public review transcripts, human touch minutes, evidence that every flagged issue is resolved, or requester outcomes. Lacking a globally locked merge queue at this volume, they use post-merge and pre-deploy tests and accept that concurrent merges may conflict, with agents to repair them. A critical new question for unresolved agreed warnings is which warning survives a merge and blocks deploy, since ‘merged’ now precedes some final integration checks. The essay does not quantify undetected conflict, rollback, outage or customer repair.

Where it sits against the other evidence. This is a deployment with first-party scale and named human choices; Chen–Stratton’s 718-firm observational study observes more proposed code and a longer ordinary PR merge queue after adoption, without a detectable Jira-resolution gain. Neither refutes the other: a bespoke internal stack at an AI lab with exceptionally high token allowance and a firm-average PR clock are different populations and denominators. Jain’s single-task harness test measures a held-out behavioral pass but no maintained deployment; Anthropic measures toolchain throughput but not independent customer acceptance. Willison guesses from public model release dates that model upgrades explain the merger surge; Anthropic says early internal Mythos Preview and Tag helped accelerate Scion and amplified demand, but neither makes a causal estimate from public release markers. Anthropic cautions Scion is not a general Git replacement: it requires a custom review/search/CI ecosystem and uniquely high model spending. Retain this as a weekend technical read for the merge/review→pre-deploy thread, not an invitation for a normal team to rewrite version control.