Cat Wu departs Anthropic after two years overseeing the development of Claude Code, the artificial intelligence company’s flagship programming tool, she announced on 3 November 2026. Her departure concludes a defining stretch in San Francisco where Claude Code evolved from an early internal research prototype into an autonomous developer agent with default permissions.
Key points
- Cat Wu stepped down as head of product for Claude Code, Cowork and Tag on 3 November 2026 after two years at Anthropic.
- Wu joined the company in August 2024 and guided Claude Code from early experiments on Claude 3.5 Sonnet to advanced workflows running on Opus 4.6.
- Her departure follows the rollout of auto mode as default, which replaced manual step-by-step approvals with automated safety classifiers.
- The transition reflects a broader shift across the industry toward autonomous AI software engineering tools.
From research product manager to Claude Code chief
Wu announced her departure on X, thanking the teams behind Claude Code, Cowork and Tag. She described the tools as the outcome of thousands of ideas from colleagues and said user feedback shaped much of what the team shipped. Wu called her time helping build Claude Code into an industry-defining product “the journey of a lifetime”.
Her path to leading Claude Code began in August 2024, when she joined Anthropic as a product manager on the Research PM team. In that role, Wu worked directly with researchers and enterprise clients to connect frontier models to practical user requirements. Before joining Anthropic, Wu worked as a product engineer at Scale AI and Dagster, followed by a period in venture capital where she wrote software to track corporate disclosures and monitor momentum across open-source code repositories.
When Anthropic opened Claude Code for internal staff use in late 2024, Wu used the early command-line interface to build internal applications. According to Wu’s account of her product management work, she spent hundreds of hours prompting Claude 3.5 Sonnet without writing code manually. She used the model to assemble Streamlit interfaces to process large volumes of user feedback, run model evaluation suites and build reinforcement-learning training environments.
Tracking model capability through Excalidraw
Wu measured the progression of Anthropic’s underlying models by running a single repeated test: asking the model to implement a functional table tool inside the open-source drawing application Excalidraw.
When she began testing the prompt against an updated version of Claude 3.5 Sonnet in October 2024, the model failed completely. By June 2025, Opus 4 managed to complete the task intermittently. By the time Anthropic introduced Opus 4.6, the model could execute the feature request reliably enough for Wu to run live demonstrations before audiences of professional developers.
That leap in raw capability changed how Anthropic structured product development. Internal roles blurred as product managers created their own prototypes and evaluation suites, designers committed code directly, and engineers took on core product decisions. Traditional product requirement documents gave way to brief, self-directed experiments and functional prototypes. That internal workflow produced several core capabilities, including Claude Code on Desktop, the AskUserQuestion interface tool and automated task lists.
The autonomy shift: from prompts to auto mode
The biggest change under Wu’s leadership was the migration from interactive prompting to autonomous execution. Early versions of Claude Code operated primarily as an assistant that answered programming questions, drafted tests and required users to manually confirm every proposed file change or command-line execution.
Wu left behind a system where Claude Code acts autonomously by default. Under Anthropic’s auto-mode announcement published in August 2026 and updated in October 2026, the tool runs in auto mode for all new sessions across Pro, Max and Team subscription tiers.
Auto mode relies on a secondary safety classifier instead of asking for manual approval on every terminal command. The classifier examines tool calls in real time, permitting safe actions to proceed automatically while halting commands that appear destructive, irreversible or directed outside the user’s workspace. If the system flags an action, Claude can either attempt a different path or ask the developer for confirmation. If the system records three consecutive blocked actions, or 20 blocks throughout a single work session, Claude Code reverts to manual step-by-step approvals.
Competition and oversight in AI development
Anthropic continues to warn developers that safety classifiers cannot eliminate execution errors. The company advises teams to maintain manual oversight whenever Claude Code interacts with production infrastructure or critical deployments. Managing safety limits has also created friction among engineers, as seen when developer pushback over safety restrictions forced discussions about false positives in automated rule-checking.
The push toward greater independence comes amid intense pressure in the AI developer market. Specialized models are challenging Claude on targeted benchmarks, highlighted by tests where automated vulnerability repair benchmarks demonstrated fierce competition from dedicated security tools. Meanwhile, rival labs are advancing their own developer agents, with DeepSeek’s prototype coding agent aiming squarely at the developer workflows Anthropic pioneered.
Wu has not yet announced her next role. Her departure leaves Anthropic to navigate the next stage of autonomous programming as coding tools move from desktop assistants to independent software agents.
