Claude Code: Behind the Scenes of an "AI-native" Engineering Organization

Anthropic Director Fiona Fung reveals how Claude Code shifted engineering to agentic workflows, achieving 100% AI-assisted commits and faster cycle times.

On stage at the Code w/ Claude conference in San Francisco, Fiona Fung, Director of Engineering for Claude Code and Claude Cowork at Anthropic, reflected on her team's transformation since agentic code became the default mode of operation. Her initial observation: writing code, tests, or refactoring no longer slows anyone down. Bottlenecks have shifted to verification, code review, and security.

Four internal standards were rewritten. Planning moves away from six-month roadmaps, which become obsolete after three, in favor of "just-in-time planning": prototyping, getting internal users on board, and iterating on their feedback. Context retrieval no longer goes through the code's author, as all pull requests are assisted by Claude, but rather through a question posed directly to the model, coupled with an automation reflex. Code review follows a "trust but verify" principle: Claude handles style, bugs, and tests, while humans focus on legal, security, and product sense. Finally, roles are blurring, PMs code, engineers take charge of design and content, and recruitment prioritizes creative profiles and systems experts over raw throughput.

All of this rests on a few non-negotiable principles: systematic dogfooding, the flattest possible team, and explicit permission to eliminate obsolete processes. Three indicators serve as a compass: onboarding time (an engineer delivers code within their first week), pull request cycle time, and the share of Claude-assisted commits, which has been 100% for the past four months, according to Fung.