Cat Wu steps down as Claude Code chief product officer after two years – Unite.AI

Cat Wu announced on

He thanked the teams behind Claude Code, Cowork and Tag, describing the products as the sum of thousands of colleagues’ ideas. He also credited user feedback with shaping much of what the teams rolled out. Wu called Claude Code’s contribution a product that defines the industry as the journey of a lifetime.

From search product management to Claude code

Wu’s March 19, 2026 account of his product management job showed how he achieved that role. She joined Anthropic in August 2024 as a product manager on the Research PM team, connecting researchers with real-world customers to improve models. Before that, he worked as a product engineer at Scale AI and Dagster, then as a venture capitalist, continuing to write code that automated company announcement tracking and identified open source projects gaining momentum.

When Claude Code became available internally in the fall of 2024, he used it to build Streamlit applications that analyzed user feedback at scale, performed evaluations to evaluate potential benchmarks, and created reinforcement learning environments to better understand training. Wu reported that these projects took hundreds of hours of prompting with Sonnet 3.5 (new), without her writing any code by hand.

He also tested subsequent models with the same exercise: adding a table tool to Excalidraw. Attempts began with Sonnet 3.5 (new) in October 2024, when Claude again failed. Wu reported that Opus 4 was occasionally successful in June 2025, and that Opus 4.6 subsequently completed those feature requests reliably enough for live demonstrations in front of thousands of professional developers.

His account linked this progress to changes in the team’s working methods. Designers shipped code, engineers made product decisions, and product managers built prototypes and evaluations, while short self-directed experiments and demonstrations replaced longer cycles of documentation. Claude Code on Desktop, the AskUserQuestion tool, and to-do lists were among the features that emerged from these experiments.

The change in autonomy he leaves behind

Wu’s departure contrasted Claude Code’s initial reliance on manually approved permission requests and uses such as answering questions and writing tests with the current default she described: trusting Claude’s judgment for self-approval, conducting large migrations, and creating new products. Anthropic’s Auto Mode announcement, published on August 7, 2026 and updated on October 22, 2026, documents the product mechanics behind this change.

The initial announcement set August 14, 2026 for the new sessions on Pro, Max and Team which were expected to run in automatic mode. The October update stated that the default would be rolling out over the following week for new sessions in most Claude apps on those plans. Instead of asking users to approve each action, automatic mode routes tool calls through a classifier intended to block irreversible or destructive actions and actions aimed outside the user’s environment, while allowing safe actions to continue.

When the classifier blocks an action, Claude can look for a safer route or ask for approval. Three consecutive blocks, or 20 in a session, brings Claude Code back to manual approvals. Anthropic warns that rating systems do not eliminate risks and recommends that users personally review Claude’s actions for any high-risk changes to production infrastructure.

For Claude Code users, the practical change is a redistribution of supervision: routine actions can move forward with fewer interruptions, while high-risk production changes still require direct human review. This balance between autonomous execution and meaningful oversight remains central to the product Wu leaves behind.