What happens after coding is solved? | Fiona Fung (Manager of the Claude Code and Cowork Teams)

Fiona Fung frames the current change in software engineering as a shift in the constraint on output. Lenny Rachitsky says Anthropic engineers produce about eight times as much code per quarter as they did in 2025, while coding itself is no longer the bottleneck. The resulting question is less whether a feature can be built and more how ambitious a team can be. Fung adds a human concern to that productivity story: engineers who work primarily with agents can lose some of the social experience of building together.

The transition is uneven. Some people are embracing AI tools and thriving, while others are frustrated or resisting them. Fung sees fear underneath some of that frustration and recommends converting it into an actionable question: what can a person do, and what remains within their control? Her advice is a stance toward a fast-moving environment, not a claim that every concern has already been solved.

Fung brings a long historical view to the discussion. She has worked as an engineer for more than 25 years, leading teams at Microsoft, Facebook, Meta, and Instagram before joining Anthropic. Her path included building TypeScript and Visual Studio, starting Facebook Marketplace, working on smart-glasses and augmented-reality efforts, and overseeing an organization of more than 500 people. That range gives her a basis for comparing the current transition with earlier changes in tools and distribution.

At IBM, Fung mainly used Vim and terminal debugging and initially did not know what an IDE was. At Microsoft, Visual Studio introduced her to integrated debuggers, breakpoints, and multithreaded debugging. She describes this as a sequence of major shifts rather than a single break with the past. Working on the Visual Studio editor also taught her the value of using the product she was building: the team received rapid feedback because its members were heavy users themselves.

Earlier software work also had a different planning constraint. When Visual Studio shipped on CDs, manufacturing deadlines made engineering time scarce and release dates hard. Online distribution changed that cadence. Fung argues that AI coding tools create another shift: more disciplines, including designers and product managers, can now submit code, and the resulting throughput makes verification a central management problem.

Fung describes an AI-first team as one in which roles move toward building. She runs a Claude Code remote session across the repositories, with access to the team's Slack channels and tracked metrics. Each month, she uses that shared context to review focus areas, shipped products, outcomes, and feedback channels with the people she supports. The session is therefore not only a code-generation tool; it becomes a management mechanism for connecting shipped work to impact.

The review goes beyond whether something shipped. Fung wants the team to examine market performance, bugs, and incidents, and to use those observations to identify themes and future investment areas. She accepts that teams will make mistakes, but wants them to learn rather than optimize for an impossible zero-error target. She says Claude has made this synthesis easier than it was when fewer things were being shipped and the review could be handled with a short manual list.

Feedback is distributed across many channels, including internal sources, email, social platforms, and messages from friends or partners. Fung used to inspect those channels during a morning ritual and decide where she could help. Routines now automate part of that process by watching a feedback channel, summarizing themes, and preparing pull requests for review. The benefit is a higher-level workflow, but the underlying requirement remains human judgment about what deserves attention.