When the tooling is the bottleneck
Al Sene, Founder & CTO
- ai-transformation
- engineering
- agentic-ai
- playbook
- productivity
When the tooling is the bottleneck
Software is changing faster than the systems that sustain it. That sentence used to be a conference slide. In 2026 it's an ops ticket.
Anthropic's engineering team just published how they rebuilt source control for that reality. In May they averaged about 2,500 merges a day. By September it was over 30,000. Git mirrors, seeds, and partial clones bought time, then stopped scaling. So a team of seven shipped Scion — an in-house, always-online source control system — migrated the monorepo off git in five months, and built the review and search surfaces around it in the same window.
Claude both created the pressure and wrote a large share of the fix. That loop is the story. The lesson for everyone else isn't "replace git next quarter." It's that responsible AI adoption multiplies productivity, and the stack that used to be "good enough for humans" becomes the bottleneck the moment the humans aren't the ones producing most of the change.
What broke wasn't "git bad"
Git is still the right tool for most repos. Anthropic is careful about that. Their history of main alone is nearly 100 GB. They do on the order of 25 million checkouts a day for CI. Fetched naively, that would be hundreds of petabytes of traffic. At peak they ran several hundred internal git mirrors.
The design assumption underneath git — every client is a full peer that can work offline — is what stopped fitting. Nearly all of their clients are agents and jobs that are online by definition. The properties that make git excellent for distributed human work were actively undermining scale.
So they flipped the principles: clients always online, server does the hard work, commits cheap and lock-free, working copy optional and sparse, everything horizontally scalable. Prior art from Google's Piper and Meta's Sapling/Mononoke mattered. The rewrite still took months of focused engineering, not a weekend prompt.
If your monorepo isn't that large and your merge rate isn't that high, you don't need Scion. You do need the question they asked: which of our systems were designed for human-scale work, and which of those will fold when AI multiplies the volume?
The operator lesson sits under the architecture
A few lines from their post should stick with anyone running an AI adoption program — not just platform teams.
Engineers didn't become less valuable. The job changed. Their team rates Claude as better at writing code than they are. What they still own is the inter-system design, the scale targets, and — often — telling the agent no. Agents will violate core constraints to satisfy a pattern that should never have been allowed. Reviewers now examine intent more than every line.
Quality went up because volume made better testing affordable. Hermetic simulations, fuzzing, rate limiting, and optimizations that would have taken experts days showed up because Claude could keep producing them on a tight timeline.
Human-scale process assumptions die quietly. A global merge-queue lock doesn't mean much when hundreds of changes can land while tests are still running. They invested in post-merge and pre-deploy automation instead. Concurrent conflicts happen; Claude often resolves them faster than a human can triage the queue.
Ops can get a new kind of attention. They put Claude "superfans" on Slack channels for individual components — obsessively watching telemetry the way a senior engineer would if they had nothing else to do. Alerts still matter. Novel, subtle failures need something that isn't one-dimensional.
None of that is "AI replaces the team." It's multiplied productivity with humans holding the design and the gates — which is exactly the posture we argue for in the 231 Labs playbook.
Tie this to how you adopt AI, not just how you ship code
Most organizations reading Anthropic's post won't rebuild source control. Many will hit a smaller version of the same pattern: coding agents clear the "write the change" bottleneck, and suddenly review queues, CI minutes, ticket hygiene, docs, and change-management rituals become the constraint.
That's where high-ROI, responsible adoption actually lives.
- Name the bottleneck you're buying. If you buy coding agents and leave review, CI, and release unchanged, you've moved the pile, not cleared it. Measure merges, cycle time, revert rate, and review edit rate — not just "lines generated."
- Keep Phase 2 (Gated Agents) honest. Soft gates fail when volume rises. We wrote about that on the approve button. Anthropic's own note — reviewers focus on intent, adversarial review can over-steer agents into complexity — is the same failure mode at platform scale.
- Treat platform and process as part of the initiative. A maturity path that only tracks model prompts and ignores the systems that absorb the output will show a fake ROI. The playbook's phases exist so leadership can see when Fluency is done and the next constraint is operational, not "more model."
- Don't copy the extreme case. Anthropic says Scion isn't a general git replacement, and their internal token spend isn't a template for every business. Steal the questions and the principles. Leave their monorepo problems until you have them.
Responsible adoption isn't slower adoption. It's adoption that compounds: more useful output per engineer, without burning trust, safety, or the platforms that have to absorb the work.
What to do this week
Pick one system that sits downstream of your coding agents — review queue, CI, staging deploys, or docs. Ask: was this designed for human-scale change volume? What breaks first if merge or ticket volume doubles? Write the answer down with an owner. That's the real backlog item Anthropic's post is pointing at for everyone who isn't running a 100 GB monorepo.
Get the playbook
The phases, gate checklist, and how we think about productivity that compounds instead of spike-and-crash are in the free 231 Labs AI Transformation Playbook at 231labs.com/playbook.