Claude Code is now authoring roughly 4% of all commits on GitHub. That is not a rounding error — it is a structural shift in how software gets written.
Over the last few weeks we dug into what agentic coding actually looks like in production across multiple client teams: where it works, where teams are getting burned, and which workflows are quietly becoming the new default in 2026.
The headline stat
4% of commits sounds small until you remember that a year ago it was essentially zero. The adoption curve is steeper than anything we have seen since the original GitHub Copilot launch.
What surprised us
- PR velocity is not the real win. Review quality is. Teams that treat Claude Code as a junior dev who needs feedback ship better code, not just faster code.
- The biggest gains are not from the heaviest users. The teams winning are the ones with the tightest guardrails around review, testing, and spec-first prompting.
- Guardrails are the moat, not the model. Anyone can point an agent at a repo. Very few teams have figured out the review loop that makes it safe.
Where it genuinely saves time
- Boilerplate migrations (framework upgrades, test scaffolding, API client generation)
- Reading and summarizing unfamiliar codebases before a refactor
- Writing tests for existing code with clear contracts
Where it creates review debt
- Large cross-cutting refactors without a spec
- Anything involving subtle business logic the agent cannot infer from context
- "Just let it run" workflows with no human checkpoint
The 2026 default
Spec-first prompting is becoming the norm. The teams that win treat the prompt like a design doc: acceptance criteria, constraints, non-goals, and a plan the agent has to confirm before touching code.
Originally published on the Horizon Tech Blog.
Top comments (0)