How Claude Code’s Team Went Four Months Without a Single Human‑Written Line of Code

AI‑Native Engineering at Claude Code

At the Code w/ Claude SF 2026 event, Anthropic engineering director Fiona Fung described how her team operated for four months without a single code commit that was not assisted by Claude, illustrating a complete shift to AI‑driven development.

Bottleneck Shift

Fiona asserted that writing code, tests, and refactoring are no longer the team’s bottleneck; the speed of code generation now exceeds the capacity of downstream processes, moving the constraint to verification, code review, and security.

Just‑In‑Time Planning

Traditional six‑month roadmaps became ineffective, so the team adopted JIT (Just‑In‑Time) planning: rapidly prototype, show work to internal users, collect feedback, and iterate, shortening planning granularity and focusing on "quick try, quick see, quick adjust" rather than long‑term design.

Changed Information Retrieval

When encountering unfamiliar code, engineers now ask Claude first, using AI to synthesize context and logic instead of locating a human code owner; this accelerates onboarding, allowing new hires to submit real code in their first week.

Layered Code Review

Claude handles style issues, obvious bugs, and test‑coverage gaps, while human reviewers concentrate on legal compliance, security, and architectural decisions, shifting human attention from checking every line to supervising AI‑handled parts.

Relaxed Role Boundaries

Product managers begin writing code and engineers engage in design work; as AI assumes many execution tasks, each role’s core value moves upward, enabling PMs to validate ideas directly and engineers to influence product decisions.

Three Non‑Negotiable Principles

Continuously use the team’s own product to surface problems quickly.

Maintain a flat structure where managers act as individual contributors to stay grounded in frontline experience.

Proactively eliminate obsolete processes, exemplified by canceling a weekly meeting that no longer added value.

Metrics for Success

Onboarding acceleration: ability of engineers to submit real code in the first week.

PR cycle time: measuring whether verification keeps pace with code generation.

Claude‑assisted commit ratio: 100% of commits in the past four months involved Claude.

Hiring Criteria in an AI‑First Environment

Fiona looks for two types of talent: product‑mindset creators with curiosity who solve real problems, and deep system experts with large‑scale distributed experience; raw throughput is less important than identifying areas still needing human judgment.

Conclusion

The transformation described is not merely a tool upgrade but a rewrite of engineering organization logic; as code generation ceases to be a bottleneck, teams must redesign planning, review, and role definitions to stay effective.