Anthropic Labs: 20-member team, biweekly reviews, 20–30% success rate

A team of about 20 people developed Claude Code, MCP, and Claude Design.

But according to Ben Mann, co-founder of Anthropic and head of Labs, most of the team’s attempts were meant to be allowed to fail.

According to Business Insider, Anthropic’s internal Labs team reviews its product prototypes on a roughly two-week cycle: to continue advancing, or to pivot? Ideas with little potential are terminated or integrated into other projects. Successful products “graduate” from Labs and are handed over to independent teams for further development.

Mann’s ideas had a success rate of only 20% to 30%. Some explorations that didn’t fully materialize were still absorbed into other prototypes or products. When viewed alongside Claude Code’s success, this figure highlights one of Labs’ most defining characteristics: embracing the reality that most attempts won’t reach completion, and reserving resources for the few directions worth continuing.

As Anthropic advances its commercialization, the importance of this mechanism is growing. According to Business Insider, the company is preparing for an IPO. Turning the capabilities of cutting-edge models into products users will consistently use remains a question Labs must continue to address.

In the interview, Mann recalled that several years ago, he had to convince the company to begin releasing products. At the time, one issue under discussion was whether the company could rely on charitable funding to sustain advanced AI research.

In 2024, he founded Labs to create space for freely exploring new product ideas. Since then, the team has incubated projects including Claude Code, the Model Context Protocol (MCP) that connects AI agents with external data and tools, and Claude Design, launched this year. Contributors include Mike Krieger, co-founder of Instagram.

Ben Mann

Mann described Labs as a true "startup incubator."

This organizational model has clear precedents. He mentioned Bell Labs, which also participated in Google’s internal incubation program, Area 120, in 2018. Google’s X also offers a similar approach: using dedicated teams to explore high-uncertainty directions, then allowing mature projects to gradually spin off independently.

What sets Labs apart is its proximity to model research. The capabilities of cutting-edge models are constantly evolving. A product that isn’t quite usable today may cross the threshold of practicality after the next model upgrade. If the product team can anticipate which capabilities are improving, they gain the opportunity to identify earlier what’s worth pursuing and what can be tested.

Claude Code was born under these very conditions.

According to Mann, researchers had already informed his team before the development began at the end of 2024 that the new model showed promise in agent programming. This sent a key signal: building products around programming could allow the model to take on more practical tasks.

Mann handed this direction to Boris Cherny, who had just joined the company. Cherny initially proposed a code analysis tool, but Mann felt the idea wasn’t ambitious enough. In an interview, he said new employees especially need to significantly raise their goals, as there are still many unknowns about what agents can ultimately achieve.

Boris Cherny

This reflects a real challenge in AI product development: if products are always designed based on already familiar capabilities, you may underestimate the tasks the model can handle next.

The prototype that Cherny subsequently developed was well received internally. In February 2025, Anthropic launched Claude Code as a preview. It leverages the underlying model to write, edit, and run code, with subsequent model upgrades continuing to drive product advancement.

Claude Code has become a major driver of Anthropic's growth. Cherny now oversees this product.

Looking back on this process, Labs’ advantage did not come from a single isolated idea. The research team first identified changes in capabilities, the product team then built prototypes based on those insights, internally tested to validate demand, and subsequent model improvements further enhanced the experience. The tighter the connection between each step, the greater the opportunity for the product to promptly capitalize on technological advancements.

However, anticipating capabilities doesn't mean every idea is worth long-term investment.

Labs conducts a "persist or pivot" evaluation for projects approximately every two weeks. This cycle is used to review direction and does not imply that all products must be completed within two weeks.

If an idea performs poorly, the team can terminate it or incorporate its valuable elements into other initiatives. Participants then move on to new tasks. Mann noted that some attempts were eventually integrated into products such as Claude’s Chrome extension.

Therefore, a success rate of 20% to 30% should not be simply interpreted as all remaining work being wasted. Even if a prototype does not evolve into a standalone product, it may still leave behind reusable features, technical expertise, or insights into user needs.

Another key rule: Projects typically graduate from Labs when their team size exceeds four members.

Claude Code and Claude Design were both transitioned into independent internal teams within the company. This allows Labs to maintain a small size and continue exploring the next set of directions. It also explains why it has always been a team with high turnover: as projects mature, some members move on, and Labs brings in new members to begin fresh initiatives.

Verified projects require greater engineering investment and long-term maintenance, while smaller teams need room to adjust their direction. Allowing mature projects to become independent in a timely manner helps prevent Labs from being overwhelmed by day-to-day operations.

It also incurs corresponding management costs. Personnel turnover is constant, requiring new members to quickly grasp research progress, while the team must continually rebuild collaborative relationships. The effectiveness of biweekly evaluations depends on the quality of judgments regarding prototype performance.

According to Mann’s description, communication between Labs and the research team is bidirectional. Product developers use new models to identify opportunities, while gaps in real-world tasks drive researchers to focus on new capabilities.

Mann summarized one of Labs’ goals as expanding AI’s “action space”—enabling powerful models to do more in the real world. This framing grounds product objectives in concrete tasks: programming requires models to manipulate code, design demands models generate more appropriate visual outputs, and cross-language use calls for systems that better understand diverse expressions. Whether the product succeeds will reveal the capabilities the model still needs to develop.

As these products enter more markets, Anthropic will also face more complex business relationships.

Forrester analyst Mike Gualtieri told Business Insider that Anthropic’s tools may directly compete with software already in use by enterprise customers. Claude Design is one example, entering the design software space occupied by companies like Adobe and Figma.

There is an ongoing tension here: Anthropic provides model capabilities to developers and enterprises while also building its own products based on those capabilities. The more features the platform adds, the greater the chance of overlap with other software products.

For users, more built-in capabilities may reduce the need to switch tools. For ecosystem partners, it is necessary to assess what differentiated value their own products can still offer. How Anthropic balances these two sides will impact the attractiveness of its platform.

If the company completes an IPO in the future, Labs will still need to maintain space for high-uncertainty exploration under more defined operational goals. This article alone cannot be used to determine the timing of the IPO, nor can the post-IPO adjustments to R&D investment be predetermined.

At least in Mann’s vision, Labs’ direction does not stop at office and development tools.

He hopes the team will eventually contribute to breakthroughs in biological research, clean energy storage, and other real-world applications. He believes accelerating fundamental scientific research will help AI deliver tangible benefits to more people.

These are still visions, with a long way to go before verifiable results. But they build on the approach Labs has already adopted: observing the boundaries of capability, identifying problems worth solving, and testing ideas through prototypes.

Claude Code has demonstrated that one of the paths is viable. The next challenge for Labs is to continuously identify the next worthwhile task amid evolving model capabilities and changing user needs.

For this team of about 20 people, the key has never been about making every project succeed. What determines how far this incubation model can go is whether it can promptly halt efforts with limited potential, allowing promising products to receive adequate resources.

Original link: https://www.businessinsider.com/anthropic-labs-team-ai-innovation-ipo-2026-9

This article is from the WeChat public account "Machine Heart".