A team of about 20 people created Claude Code, MCP and Claude Design.
But Ben Mann, co-founder of Anthropic and head of Labs, believes that most of the team's attempts should have been allowed to fail.
According to Business Insider, the Labs team within Anthropic reviewed the prototypes of the product in hand for a period of about two weeks: continuing to advance or adjusting direction? The idea of a lack of prospects would be terminated, or incorporated into other projects.
Mann's idea of a 20% to 30% success rate. Some unsuccessful explorations can also be absorbed into other prototypes or products. This number, combined with the success of Claude Code, forms the most striking feature of Labs: accepting a large number of attempts without reaching the end, leaving the resources to a few that are worth continuing to invest in.
The importance of this mechanism is increasing as Anthropic advances commercialization. According to Business Insider, the company is preparing to go public. How to turn front-end model capabilities into products that users are willing to continue to use will be a question that Labs will need to continue to answer.
How labs work like a startup incubator
Mann recalled in an interview that a few years ago, he even needed to convince the company to start launching products. At the time, one question that needed to be discussed was whether the company could rely on charitable funding to continue advancing advanced AI research.
In 2024, he founded Labs, a space for free exploration of new product ideas. Since then, the team has been involved in the development of Claude Code, the Model Context Protocol that connects AI with external data and tools (MCP), and Claude Design, launched this year. Instagram co-founder Mike Krieger, among others.
Ben Mann
Mann describes Labs as a truly "entrepreneurial catalyst".
This form of organization has clear references. He mentioned Bell Laboratories, which joined Google's internal optimization project Area 120 in 2018. Google's X also offered an idea: to explore higher uncertainty directions with a dedicated team, and to make mature projects gradually independent.
What is special about Labs is that it is more reflected in the distance between it and model research. Continuous changes in the capabilities of front-end models. A product that is not useful enough today may cross a practical threshold after the next round of model upgrades. If the product team knows in advance which capabilities are improving, there is an opportunity to judge earlier, what is worth doing, what can be tried.
Claude Code was born under these conditions.
The birth of Claude Code
According to Mann, even before the start of development at the end of 2024, the researchers revealed to his team that the new model showed potential for intelligent programming. This gave an important signal: around programming building products, attempts could be made to make the model take on more practical operations.
Mann gave this direction to Boris Cherny, who had just joined the company. Cherny initially proposed a code analysis tool, but Mann felt that the idea was not large enough.
Boris Cherny
This phrase corresponds to a real difficulty in AI product development: if the product is always designed according to familiar capabilities, it is possible to underestimate what the model can do next.
Cherny's subsequent prototyping was well-received internally. In February 2025, Anthropic released Claude Code as a preview. It used underlying models to write, edit and run code, and subsequent model upgrades continued to drive product development.
Claude Code became an important driver of Anthropic's growth. Cherny is now responsible for the product.
Failure of the product mechanism
Looking back at this process, the advantages of Labs do not come from an isolated idea. The research team first noticed a change in capabilities, the product team built prototypes accordingly, used internal verification needs, and subsequent model advances continued to improve the experience.
However, seeing the potential beforehand does not mean that every idea is worth investing in in the long run.
Labs conduct a "persevere or reverse" evaluation of the project approximately every two weeks. This cycle is used to check directions and does not mean that all products must be completed within two weeks.
If an idea doesn't work well, the team can terminate it or incorporate valuable parts of it into other explorations. Participants then move on to new tasks. Mann notes that some attempts were eventually integrated into Claude's Chrome expansions and other products.
Consequently, a 20% to 30% success rate cannot simply be understood as the total scrapping of the remaining work. A prototype that does not grow as a standalone product may still retain reusable functionality, technical experience or needs judgement.
Another rule is equally critical: When a project team is typically larger than four people, it is "graduated" from Labs.
In this way, both Claude Code and Claude Design transitioned into independent teams within the company. Labs thus maintained a smaller scale and continued to explore the next batch of directions. This also explains why it has always been a more frequently staffed team. After the project matured, some members left; Labs added new members and started new trials.
Bilateral feedback between research and products
Projects that have been validated require more engineering input and long-term maintenance, while smaller teams need to keep the space to adjust directions. Making mature projects independent in a timely manner will help prevent the labs from being fully occupied by day-to-day operations.
It also has corresponding management costs. Personnel is constantly changing, new members need to quickly understand the progress of research, and teams need to constantly rebuild collaborative relationships. Whether a bi-weekly assessment is effective depends on the quality of the judgement on the prototype performance.
From Mann's description, the interaction between Labs and the research team is also two-way. Product developers use new models to look for opportunities, and gaps in actual tasks will encourage researchers to focus on new capabilities.
Mann generalizes one of the goals of Labs as expanding AI's "space of action", that is, enabling powerful models to do more in the real world. This expression places the goal of product work on specific tasks. Programming requires models to be able to operate code, design requires models to generate more appropriate visual results, and cross-language use requires systems to understand more people's expressions.
Ecological challenges after commercialization
As these products enter more markets, Anthropic will also face more complex business relationships.
Forrester analyst Mike Gualtieri told Business Insider that the tools launched by Anthropic could compete directly with software already used by enterprise customers. Claude Design is an example of how it has entered the design software space of companies such as Adobe, Figma and others.
There is an ongoing tension: Anthropic provides both developers and enterprises with modeling capabilities and builds their own products based on these capabilities. The more functionality the platform increases, the greater the chances of overlapping with other software products.
For users, more built-in capabilities may reduce tool switching. For eco-partners, it is necessary to judge what differential value their products can offer. How Anthropic balances the two ends will affect the appeal of its platform.
If the company were to go public in the future, Labs would still need to reserve space under clearer operational targets to explore for greater uncertainty. It is not possible to judge its time on the market based on this report alone, nor is it possible to determine in advance how research and development investment will be adjusted after it goes public.
At least in Mann's conception, the direction of Labs did not stop at the office and development tools.
He hopes the team will be able to participate in future breakthroughs in bio-research, clean energy storage and other real-world applications. He believes that accelerating basic science research will help AI bring real benefits to more people.
These are still visions, and there is still a long way to go from verifiable results. But they continue the way the Labs have been working: observing the limits of capability, finding problems worth solving, and then prototyping ideas.
Claude Code demonstrates that one of these paths can be followed. Labs' next test is to continually find the next task worth investing in between changing model capabilities and user needs.
For this team of about 20 people, the key may never have been to make each project a success. Whether to stop the limited prospects in time and get enough resources for a potential product will determine how far this mechanism can go.
