A team of about 20 people created Claude Code, MCP, and Claude Design.
But in the eyes of Anthropic co-founder and Labs head Ben Mann, most of this team's attempts should have been allowed to fail from the start.
According to Business Insider, the Labs team within Anthropic operates on roughly two-week cycles to review product prototypes at hand: move forward, or pivot? Ideas lacking promise are terminated or folded into other projects. Products that truly gain traction 'graduate' from Labs and are handed off to independent teams for further development.
Mann cites an idea success rate of only 20% to 30%. Some explorations that don't become standalone products are also absorbed into other prototypes or products. This number, placed alongside the success of Claude Code, defines the most distinctive feature of Labs: accepting that many attempts won't make it to the finish line, reserving resources for the few directions worthy of continued investment.
As Anthropic pushes forward with commercialization, the importance of this mechanism is rising. According to the Business Insider report, the company is preparing for an IPO. How to turn cutting-edge model capabilities into products that users are willing to use consistently will be a question Labs needs to keep answering.
How Labs Operates Like a Startup Incubator
In an interview, Mann recalled that a few years ago, he even needed to convince the company to start releasing products. At the time, one question for discussion was: could the company rely on philanthropic funding to continue advancing AI research.
In 2024, he created Labs, carving out space for the free exploration of new product ideas. Since then, this team has incubated Claude Code, the Model Context Protocol (MCP) connecting AI agents with external data and tools, and this year's Claude Design. Instagram co-founder Mike Krieger, among others, has been involved.
Ben Mann
Mann describes Labs as a 'startup incubator' in the true sense.
This organizational form has clear precedents. He mentioned Bell Labs and also joined Google's internal incubator program Area 120 in 2018. Google's X offers another reference point: using dedicated teams to explore directions with higher uncertainty, then letting mature projects gradually become independent.
The distinctive feature of Labs lies more in its proximity to model research. The capabilities of frontier models are constantly changing. A product that isn't quite usable today might cross the usability threshold after the next model upgrade. If product teams can know in advance which capabilities are being improved, they have a chance to judge earlier what's worth doing and what to start testing.
Claude Code was born under precisely such conditions.
The Birth of Claude Code
According to Mann, even before starting development around late 2024, researchers had informed his team that the new model showed potential in agentic programming. This gave a crucial signal: building a product around programming could involve trying to let the model take on more operational tasks.
Mann gave 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 the interview, he said new employees especially need to raise their ambitions significantly, because there's still much unknown about what agents can truly accomplish.
Boris Cherny
This statement points to a practical challenge in AI product development: if one always designs products based on currently familiar capabilities, it's easy to underestimate the tasks the model might be able to handle next.
The prototype Cherny subsequently built gained internal popularity. In February 2025, Anthropic launched Claude Code as a preview. It uses the underlying model to write, edit, and run code, with subsequent model upgrades continuously pushing the product forward.
Claude Code became a major driver of Anthropic's growth. Cherny now leads this product.
A Product Mechanism That Allows Failure
Looking back at this process, the advantage of Labs didn't come from a single isolated idea. The research team first identified capability changes, the product team built prototypes accordingly, internal use validated demand, and subsequent model improvements continued to enhance the experience. The tighter the connection between these steps, the better the product's chance of catching technological progress in time.
However, seeing capabilities early doesn't mean every idea deserves long-term investment.
Labs conducts a 'persevere or pivot' review of projects roughly every two weeks. This cycle is for checking direction, not that all products must be completed within two weeks.
If an idea isn't performing well, the team can terminate it or fold its valuable parts into other explorations. Participants then move on to new tasks. Mann mentioned that some attempts were eventually integrated into products like the Claude Chrome extension.
Therefore, the 20% to 30% success rate can't be simply understood as the remaining work being completely scrapped. A prototype that doesn't grow into a standalone product can still leave behind reusable features, technical experience, or demand insights.
Another rule is equally crucial: When a project team typically grows beyond four people, it 'graduates' from Labs.
Both Claude Code and Claude Design transitioned this way into independent teams within the company. This keeps Labs relatively small, continuing to explore the next batch of directions. It also explains why it's always a team with relatively high personnel turnover. As projects mature, some members leave with them; Labs then adds new members to start new attempts.
Two-Way Feedback Between Research and Product
Validated projects need more engineering investment and long-term maintenance, while small teams need to retain room to pivot. Allowing mature projects to become independent in time helps prevent Labs from being completely consumed by daily operations.
It also has corresponding management costs. With personnel constantly changing, new members need to quickly understand research progress, and the team must continually rebuild working relationships. The effectiveness of the bi-weekly reviews depends on the quality of judgment regarding prototype performance.
From Mann's description, communication between Labs and research teams is also bidirectional. Product developers leverage new models to find opportunities, while gaps in real-world tasks prompt researchers to focus on new capabilities.
Mann frames one goal of Labs as expanding AI's 'action space,' meaning enabling powerful models to do more things in the real world. This formulation grounds the goal of product work in specific tasks. Programming requires models to manipulate code, design requires models to generate more appropriate visual outputs, and cross-language use requires systems to understand more people's expressions. Whether a product can succeed exposes the capabilities the model still needs to acquire.
Ecosystem Challenges Post-Commercialization
As these products enter more markets, Anthropic will also face more complex commercial relationships.
Forrester analyst Mike Gualtieri pointed out to Business Insider that the tools Anthropic launches might directly compete with software already used by enterprise customers. Claude Design is one example, entering the design software field occupied by companies like Adobe and Figma.
A constant tension exists here: Anthropic both provides model capabilities to developers and enterprises and builds its own products based on these capabilities. The more features the platform adds, the more opportunities for overlap with other software products arise.
For users, more built-in capabilities might reduce tool switching. For ecosystem partners, they need to assess what differentiated value their own products can still provide. How Anthropic balances these ends will influence its platform's appeal.
If the company completes an IPO in the future, Labs will also need to preserve space for uncertain explorations under clearer business objectives. We cannot judge its IPO timeline based solely on this report, nor predetermine how R&D investments might adjust post-listing.
At least in Mann's vision, Labs' direction doesn't stop at office and development tools.
He hopes the team can participate in biological research, clean energy storage, and breakthroughs in other real-world applications in the future. He believes accelerating fundamental scientific research will help AI deliver tangible benefits to more people.
These remain aspirations, a long way from verifiable outcomes. But they continue the working method Labs has already adopted: observing the boundaries of capabilities, finding problems worth solving, and then testing ideas through prototypes.
Claude Code proved that one path can work. The next test for Labs is to continuously find the next task worth investing in, amidst ever-changing model capabilities and user needs.
For this team of about 20 people, the key perhaps never lies in making every project succeed. Whether they can promptly stop attempts with limited prospects and provide sufficient resources to promising products determines how far this incubation mechanism can go.
Original link: https://www.businessinsider.com/anthropic-labs-team-ai-innovation-ipo-2026-9
This article is from the WeChat public account "机器之心" (The Heart of Machines)
