Anthropic says its daily merged code changes grew from roughly 2,500 in May to over 30,000 in September. Its response was Scion, an internal replacement for Git built around always-online agents. The interesting part isn’t another version-control system: it’s what happens to engineering judgment when producing code stops being the scarce resource.

  • Seven engineers built Scion, its review tool, and related systems, then migrated the company’s shared repository in five months, according to the team.
  • Agents still needed people to say no: some scanned entire change histories instead of asking the right service, while adversarial review pushed others toward unnecessary defensive complexity.
  • The team says it reviews every change request for Scion, but focuses on the change’s intent rather than reading every line.
  • With changes arriving faster than automated tests could finish, the company invested in testing after merging and before deployment rather than a globally locked merge queue.

That last tradeoff matters more than the throughput chart. Faster generation moves the question from “can we write this?” to “can we tell whether this is safe and useful?” Anthropic explicitly warns that its enormous repository and unusually generous internal model budget make this a poor general-purpose replacement for Git.

The 398-comment thread on Hacker News adds doubts about maintainability, an alternative scaling approach, and a proposal for recording what agents meant to do.

What the thread adds

  • kburman — “The tradeoff between building something custom and adapting an existing solution is changing. Whether maintaining all that custom software turns out to be equally cheap is another question.”
  • armcat — proposes tracking ““intent” (what was the agent trying to ACTUALLY accomplish), “provenance” (which agent made what decision - based on what evidence), and “continuous verification” (can correctness be maintained through ongoing validation, repair, and rollback rather than a single pre-merge testing gate).”
  • Eridrus — offers Cursor’s Git-scaling write-up as a counterexample: “Cursor scaled git to way more writes/s than Anthropic did with their rewrite.” This is the commenter’s comparison, not a verified like-for-like measurement.
  • nater5000 — on why agents still need history: “a code change seems appropriate, it is implemented, it turns out there’s some external factor which makes it not appropriate, the change is reverted, then that same code change seems appropriate again, etc.”

The unanswered question: what does the churn accomplish?

Two sampled commenters ask where the volume turns into useful work. netburst asks, “What is the code doing? What problems are being solved?” ihumanable asks, “Is there a bunch of code they are using as part of training, or as pre-processing or post-processing the requests to the model?” The article identifies research and service development as uses, but doesn’t break down the merge count by purpose or demonstrate the value of that volume.

Handles are pseudonymous; HN publishes no per-comment scores. The fetched order is HN’s ranking, while these excerpts are arranged for usefulness. This is a slice of the thread, not a consensus.