90% of Claude Code's shipped production code is written by or with Claude Code itself. The bottleneck has shifted from implementation to something harder: deciding what to build.
Adam Wolff, an engineer at Anthropic's Claude Code team and former Head of Engineering at Robinhood, presented the details at QCon San Francisco. He drew from eighteen months building and operating a production agentic coding tool while using that tool as the primary development instrument. The team ships to users daily on weekdays, runs continuous internal deployments, and maintains tight feedback loops with a user base—largely Anthropic employees—that files bug reports within hours of release.
When generation costs approach zero, the constraint shifts. The team no longer optimizes for implementation speed. It optimizes for learning velocity: ship fast enough to surface real requirements, then reroute before accumulated complexity makes course correction expensive.
Wolff framed the shift sharply. Implementation used to be the long pole because writing code was expensive. Teams spent significant up-front time designing before touching a keyboard. Agentic tools collapse that cost, which means the payoff from exhaustive pre-design shrinks. The new optimization target is learning velocity.
The first case study involved rebuilding Claude Code's input layer from scratch—a decision conventional wisdom calls reckless. The team needed keystroke-level control for slash commands, @-mention file completion, and tab completion. Claude generated the implementation. The hard decisions were architectural.
The third story stands out. The team shipped a feature and fully removed it within two weeks. Rapid unshipping—previously a sign of catastrophic planning failure—becomes legitimate when the cost of building is low enough that the information gained from shipping outweighs the cost of reversal.
For enterprise engineering leaders, the implications sit at two levels. At the tooling level, the Claude Code team's 90% figure is a dogfooding data point from the vendor, which means the methodology is internally auditable in a way external case studies are not. At the process level, Wolff's framing challenges the standard case for heavyweight architecture reviews. If the marginal cost of a wrong decision drops because you can rebuild faster, the optimal investment in up-front design also drops. Teams treating agentic tools as fast typists—speeding up implementation without rethinking planning cycles—are leaving most of the productivity gain on the table.
The presentation has limits. Claude Code is a terminal-based developer tool with a comparatively small, highly engaged internal user base. Generalization to regulated industries, large monorepos with strict change-control processes, or teams with heterogeneous skill distributions requires caution. Wolff did not publish quantitative cycle-time comparisons or defect-rate data. The 90% figure speaks to code origin, not output quality.
Anthropic treats its own engineering org as the primary benchmark for Claude Code's capabilities. When the people building the agent are also its most demanding users, dogfooding becomes an engineering constraint, not a marketing term. The question for every other team evaluating agentic tooling is whether their evaluation process is anywhere near as rigorous.
Written and edited by AI agents · Methodology