Cat Wu

Cat Wu is Anthropic’s head of product for Claude Code, shaping how AI agents write software, collaborate across teams, and operate with increasing independence. Her work treats product judgment—not implementation speed—as the emerging constraint on what software teams can accomplish.

From engineering and investing to Claude Code

Wu worked as a product engineer at Scale AI and Dagster before moving into venture capital, where she built software to monitor startup announcements and identify promising open-source projects. She joined Anthropic in August 2024, initially helping connect model research with customer needs.

An early internal version of Claude Code became her laboratory: she used it to analyze user feedback, run evaluations, and explore reinforcement-learning environments. Starting in October 2024, she repeatedly challenged successive Claude models to add a table tool to Excalidraw. Their progress—from repeated failure to reliable implementation—became a practical lesson in building products amid rapidly changing model capabilities, which she described in her first-person account of AI product management.

Her product responsibilities have expanded to include Claude Cowork and collaborative agent experiences. Wu expects product managers to prototype, engineers to develop stronger commercial and product instincts, and designers to participate more directly in implementation.

• Product taste becomes more valuable as code becomes cheaper. When working software can emerge within days instead of months, deciding what deserves to exist becomes more consequential than producing extensive specifications. Wu wants teams to test ideas directly and use adoption and retention to decide what ships.

• Multiplayer coding agents belong inside existing team workflows. With Claude Tag, colleagues can direct a shared agent in Slack, carry work between product, design, and engineering, and establish channel-specific preferences. The agent can monitor bug reports proactively and move selected issues toward pull requests.

• Autonomy depends on graduated trust and agent evaluation. Wu favors keeping human code owners responsible for sensitive changes while expanding automated review where incident-informed tests and behavioral evaluations support it. Security work must address prompt injection and data exfiltration, while safeguards such as Claude Code’s auto mode still require judgment about high-stakes actions.

• Agent instructions need context, not indiscriminate rules. Wu argues that mandates such as always verifying interface changes can misfire when applied to trivial edits. Effective agent tools should have clearly differentiated purposes and make consequential actions legible to users.

Wu has also highlighted browser access inside Claude Code and long-running autonomous work, extending coding agents beyond isolated implementation. Outside work, she has used Claude to build a climbing-project tracker and research climbing trips around routes, travel logistics, and short approaches.

In September 2026, Wu described using redesigned Projects to coordinate sessions, retain evolving memory and receive aggregated status. Anthropic’s launch documentation identifies this as a limited beta for select Pro and Max Claude Code users, with parallel cloud sessions and broader rollout to follow. It arrived alongside the initial unification of Cowork and chat, extending her work on delegating longer tasks.