Claude Code writes 4% of all public GitHub commits — what it means for teams

In March 2026 SemiAnalysis published a report: 4% of public GitHub commits came through Claude Code. The projection is 20% by December. Pragmatic Engineer surveyed 906 engineers in February — 46% picked Claude Code as their most-loved tool. The number sounds like Anthropic marketing. The method is open: they count commits with Co-Authored-By Claude or with specific author patterns. Let me share what this actually changes on the ground.

What changed inside our studio

Team of six developers. Before: junior commits, senior reviews, junior fixes. Now: junior runs Claude Code, gets a first version in minutes, review happens on working code. Cycle time from brief to staging dropped from three days to one.

There is a catch. PR size roughly doubled. We used to ship 200-400 lines a day. Now it is 500-900. Review time stayed the same because 70% of those lines are routine and easy to scan. The bottleneck shifted from writing to deciding.

What we do not delegate

Six months of mistakes gave us this list:

• Architecture decisions — Claude writes great code inside a chosen design, but picks poorly between two non-trivial options. Given a choice between a module and a service, 8 out of 10 times it goes for the heavier option
• Queries against legacy databases — especially with non-standard schemas or historical hacks. Claude reasons about how things should be, not how they are. We once shipped a migration that wiped a production cache
• Integrations with Russian payment and gov services — YuKassa, SBP, Kontur, Gosuslugi. Sparse docs, little training data. The agent confidently writes code patterned on similar foreign APIs and breaks at edge cases
• Dependency picks — Claude likes pulling in new packages. We routinely catch 50KB libs added for two lines of code in review

What happened with junior hiring

This is the most painful industry conversation of 2026. In Q1 three studios we know stopped hiring juniors entirely. Claude closes most mid-level tasks except client communication and context. We landed in the middle — we still hire juniors, but fewer of them, and we train them on agent-paired work from day one.

Interns who started with Claude Code from day one ended up stronger after six months than those who spent six months coding by hand. They iterate faster, read more code, frame problems better. There is a flip side — they struggle without the agent. When Claude is wrong and you have to go in with print statements, they freeze. We added a week to onboarding where Claude is banned, so the base skill does not atrophy.

How review changed

The old PR review scored on comment count. That does not work anymore — comments dropped because small mistakes dropped. Bigger mistakes got more dangerous: code looks right, passes tests, does the wrong thing. We added a rule: every reviewer writes "what would I have done differently from the agent." If there is nothing to write, the review is sloppy and goes back. Sounds formal. It works. Three months in, we caught two bugs that would have made it to prod otherwise.

Is the 20% projection realistic?

Yes, and probably conservative. Cursor and Codex stamp their own marks in commit metadata, so the real share of AI-generated code is already above 4%. By December most teams will face a choice: build agents into the process or fall behind on timelines for those who did. This is not a warning, it is just the next step of automation — IDEs replaced notepads, git replaced zip archives, agents are the next layer.

What I took away: the agent does not replace the engineer, it relocates the bottleneck. The constraint used to be typing speed. Now it is thinking quality. That is a useful constraint to be stuck on.