Claude Code Product Lead Cat Wu Joins OpenAI; Three Core Members Leave Anthropic Within Ten Days

People and infrastructure are in place, and OpenAI is betting on team agents.

English text is being prepared.

English text is being prepared.

This morning, former Claude Code product lead Cat Wu announced on X that she is joining OpenAI as a product lead in the core product group. A day earlier, she had just said goodbye to Anthropic—a seamless transition, so to speak. OpenAI President Greg Brockman immediately reposted a welcome.

This is already the third core member to leave the Claude Code team within ten days. Previously, the departures were Anthony Morris, the father of Claude Code's desktop version, and Thariq Shihipar, a core Agent engineer. And this timing comes exactly as Anthropic stands at a critical juncture on the road to going public.

Cat Wu's departure shook Silicon Valley not only because she switched camps, but also because she took with her the complete methodology of this most profitable product line.

01

Who is Cat Wu?

Cat Wu, whose real name is Catherine Wu, is a Chinese-American star product manager who is highly sought after in Silicon Valley's AI circles.

As the overall product lead for Anthropic's flagship AI coding tool Claude Code and its office agent Cowork, she and "father of Claude Code" Boris Cherny are known as the "Batman and Robin" of the Anthropic team, and she is one of the key figures who personally turned Claude Code from an obscure internal tool into a global hit.

She was also a top student at Princeton, where in her senior year she co-developed the collaborative whiteboard software Scratchwork at an on-campus incubator; she has honed her skills at big tech companies and unicorns, writing code and researching data; and she spent three years as a venture capitalist focused on developer tools and AI infrastructure. Her rich cross-disciplinary experience has made her the rare kind of person who can read code, calculate the business math, and accurately grasp what developers want.

In August 2024, she joined Anthropic, initially as a product manager on the research side.

At that time, Claude Code was still an inconspicuous experimental project inside Anthropic. Cat Wu wanted to see what this thing could actually do, so she had it help her build a reinforcement learning environment from scratch. She spent dozens of hours repeatedly describing requirements and adjusting direction, without writing a single line of code from start to finish. When the environment finally ran successfully, she was stunned by the experience and immediately applied to transfer into this fledgling small team.

In that team, she and technical lead Boris Cherny formed a golden partnership. Boris was responsible for writing the initial prototype code and keeping an eye on what could be achieved next; Cat Wu was responsible for turning ideas into products and mobilizing resources across the board to bring them to market. With extremely strong execution, she pioneered a very rapid product iteration model, directly compressing development cycles from months to weeks, or even counting by days.

Later, her responsibilities extended further to Claude Cowork, exploring AI in team work scenarios.

Today, Claude Code is one of the hottest AI coding agents in the world, used by a large number of developers. In February 2026, Anthropic disclosed in its Series G financing announcement that Claude Code's annualized revenue run rate (ARR) had exceeded $2.5 billion, directly doubling from the beginning of the year.

Across the entire history of technology development, this is one of the fastest-growing enterprise software (SaaS) products, and it was only opened to the public in May 2025. The product line Cat Wu oversaw is the most core growth driver of this revenue curve.

Cat Wu is an unconventional person. At Anthropic, she redefined "how product managers should work in the AI era." She publicly shared several methodologies on podcasts that are still regarded by peers as gospel:

For example, long PRDs are dead. Today, when model capabilities change every few months, if a PM has an idea, they should quickly build a working prototype themselves or have AI do it, and iterate on the product directly through running and evaluation; there is also the "three-month obsolescence principle": any product plan longer than six months will become invalid, and as long as the experience can be improved today, it should be shipped immediately. When upgrading to a new model, she even decisively cut 20% of the old prompt patches that the team had painstakingly accumulated; she encouraged engineers to take on "side quests" outside the main line. In this atmosphere that encourages free exploration, engineers spontaneously stumbled upon many classic features during daily chats and afternoon tea conversations: the now globally popular Claude Code desktop version and the highly inspired "AskUserQuestion" feature both came about this way.

In her departure post, Cat Wu expressed pride in Claude Code. The product has made a key leap: from the early days when users had to manually approve permissions one by one, to now being able to trust the model's own judgment and complete permission approvals automatically. And the insight she gained on this journey, as she previously shared on a podcast, is that the most important thing in building AI-native products is learning to let go.

02

On the eve of the IPO, is this departure timing too sensitive?

Some may ask: by leaving on the eve of Anthropic's IPO, won't she lose out on a huge sum of money? Actually, that is not the case.

Some netizens pointed out that she had been at Anthropic for about 26 months. Under Silicon Valley's common four-year option vesting rules, she had already received about half of her shares. Although before the company goes public these options are only "paper wealth," as long as the company completes its bell-ringing listing, the portion she has already vested can still be cashed out.

In other words, she has already obtained a ticket to financial freedom and no longer needs to remain "tied" to the company.

In fact, she is not the only one choosing to exit at this time. On October 24, two other core members of the team, Thariq and Anthony, also got ahead of her and announced their departures. Just two days later, Anthropic officially submitted its S-1 prospectus to the SEC, planning to launch its roadshow on November 9. The data in the prospectus is extremely impressive: revenue of more than $16.2 billion in the first half of 2026, with a valuation heading straight toward the trillion-dollar scale.

The timing of Cat Wu's departure comes right on the eve of the company's transformation from a "startup legend" into a "public company." For employees who stay, going public means a wealth carnival; but for Cat Wu, the "number one" who has just taken the product from zero to $2.5 billion in ARR (annual recurring revenue), this is precisely the peak of her professional reputation. With this dazzling track record, she can naturally and smoothly begin her next adventure.

The capital market does not care how much money she personally gave up; investors care more about the dangerous signal sent by this kind of "collective departure":

Claude Code is Anthropic's main engine for commercialization and the company's most important card in competing against OpenAI's coding products. In just ten days, the product head of this core business line, the desktop version lead, and a core Agent engineer left one after another, and the number one even jumped directly to the biggest competitor. For any company sprinting toward an IPO, this is a governance issue that investors will repeatedly weigh.

After joining OpenAI, Cat Wu will report directly to OpenAI's "Reset master" Tibo.

Tibo is no ordinary person. He was previously at Google DeepMind, and after joining OpenAI, he led the core product line all the way from early large language models to ChatGPT and then Codex. Placing the number one product lead of Claude Code under him shows that what Cat Wu is taking over is not some marginal project, but the product center of OpenAI's entire agent product line.

She has already run the full loop at Anthropic and knows how to polish model capabilities in the lab into products that developers cannot do without. And that is exactly what OpenAI lacks most right now.

Next, she will collaborate with Nick Turley, who is in charge of enterprise products, and Ashley Alexander, who is in charge of consumer products, to specifically build team-level AI assistants.

Another noteworthy matter: the day before Cat Wu joined, OpenAI announced the acquisition of Pierre Computer. Many people's first reaction was that OpenAI was building another competing product to take on GitHub, but that is not what this is at all.

Traditional Git architecture was designed from the very beginning for humans writing code. A developer commits code a dozen or so times a day, and version management and permission division are more than sufficient. But if a team has five or six frontend, backend, and testing Agents continuously writing code into the repository at the same time, the old Git architecture will quickly surrender amid repeated version conflicts and permission overreach.

Pierre solves exactly this problem. It focuses on native code storage for AI agents, specifically handling version management and concurrent read/write challenges when massive numbers of Agents write code at high speed concurrently. Its two open-source codebases, Trees and Diffs, were integrated into Codex early on. Combined with OpenAI's internally developed Git system Pando, it has pushed code push latency down to a level usable for real-time collaboration.

Putting these moves together, OpenAI's intentions can no longer be hidden. It does not want to be just "an AI that helps you write code"; it wants to connect the entire chain from ideas, writing code, storing code, and managing versions to release, turning Codex into the main platform for collaborative software development between humans and AI, directly challenging Microsoft's GitHub.

Now the industry is all saying that "team Agents are the next battlefield," but why can Agents still never step into enterprise production environments?

It is not at all that coding accuracy is insufficient; it is that in team scenarios no one dares to really use them. Do you dare let an Agent directly touch the production directory? If it accidentally changes a core module written by a colleague and commits without running tests, who is responsible for the online incident?

In what scenarios an Agent can directly modify files, what operations must stop and bring in a human for approval, and how to split tasks so that several Agents working at the same time do not fight each other—these detailed judgments are all product instincts developed through hands-on experience, and this happens to be exactly the experience Cat honed over two years at Claude Code from zero.

In the future, when an Agent finishes writing code, another Agent responsible for testing will automatically run checks and pre-checks, and humans will only need to make decisions at key nodes. Internally, OpenAI calls this new workflow natively designed for Agents "Reset." At present, it has 1 billion users, and building team-level Agents is a natural move.

By comparison, Anthropic has the best coding model but only a base of a few hundred thousand developer users; Cursor rose on developer reputation but lacks underlying model capabilities and infrastructure.

At this point, with the three puzzle pieces of people, technology, and infrastructure in place, OpenAI has truly obtained its entry ticket to the team Agent track.