Anthropic's flagship product was built by twenty people

The team has existed inside the company since 2024, but its name only became public recently. On January 13, 2026, Anthropic published a post on its company blog titled "Introducing Labs," disclosing for the first time that such a team existed and who runs it. In that post, the company said Instagram co-founder Mike Krieger had stepped down as Chief Product Officer to co-lead Labs alongside Ben Mann, while Amy Bohra would take over the rest of the product organization.

What Labs actually does is simple. The team builds a product around something Claude has only just become capable of doing, ships it to early users in a rough, unpolished state, and only turns it into a fully built-out product if it gets real traction. In announcing Labs, President Daniela Amodei said the pace at which AI is advancing demands a different way of building and organizing teams altogether.

The key detail here is approval. As companies grow, launching even a single new product typically means writing proposals and sitting through approval meetings. Labs was deliberately carved out as a space where that process doesn't apply, so the team can build first and ask questions later. As a result, some of what the team has shipped is the kind of thing that would never have cleared approval anywhere else in the company.

To put it plainly: MCP is a shared protocol that lets AI connect to a company's various internal programs and data in the same standardized way. It's a bit like having a single universal adapter instead of carving a different plug for every outlet.

Claude Code, now a major pillar of Anthropic's revenue, also started as an internal experiment. Boris Cherny built the prototype alone in September 2024, and the second engineer, Sid Bidasaria, along with product manager Cat Wu, joined in November. The project didn't start with a product team attached — it began as a tool one person built to make his own work easier.

Internal reaction became the proof of concept. On the day it was first released internally, 20% of Anthropic's engineers used it; within five days, half the company had adopted it. Since the tool had already been market-tested inside the company before ever going external, there was no need for lengthy debate over whether to turn it into a real product.

By around July 2025, the team had grown to roughly ten engineers, and it has since expanded into a full product organization spanning engineering, product, design, and data. What this team did in the stretch between one person and ten is really the core of this story.

The numbers make the pace of shipping clear. Internally, the team pushes 60 to 100 new releases a day; externally, it ships once a day. As of summer 2025, each engineer was submitting around five pull requests a day.

One striking detail: when the team doubled in size, throughput rose 67%. Normally, adding more people slows things down temporarily as new hires get up to speed — but here the opposite happened. The team credits this to tooling that cuts down the time it takes to onboard new members.

There's also a deliberate logic behind the tech choices. TypeScript and React were chosen because they're the languages the model knows best, which is why roughly 90% of Claude Code's own codebase was written by Claude Code itself. In public remarks in September, Ben Mann said Claude writes 95% of this team's code.

That doesn't mean the model gets a free hand, though. Boris Cherny has drawn a firm line: running Claude Code doesn't mean it can change a system without the user's permission. The teams handing the most work off to the model tend to be the ones that have already decided, in advance, what they won't hand off.

The detail most often missed in all this is who the very first user was. Claude Code's first market wasn't external developers — it was Anthropic's own engineers. Because the people who built it used it every day for their actual work, friction showed up not in meeting notes but in the day-to-day grind, which is how the team ended up producing more than twenty screen mockups for a single feature in just two days.

So what Labs has produced doesn't stop at Claude Code. Skills, Claude for Chrome, and Cowork, released on January 12, 2026, all came out of the same team. Cowork launched under a "Research Preview" label, signaling that it isn't finished and inviting feedback rather than presenting it as a polished product. None of the four launched as a finished product — all shipped rough and unfinished first, and only the ones that attracted real usage were eventually hardened into full products.

That label itself is a way of setting expectations up front. By telling users in advance that something isn't a finished product yet, shipping it in a rough state becomes a deliberate part of the process rather than a mistake.

The results of this approach are already visible in the numbers. Claude Code went from an initial test release to a billion-dollar product in six months, and MCP has climbed to around 100 million downloads a month. That scale of impact from a twenty-person team is strikingly out of proportion.

Labs started as a very small team in 2024 and now plans to double its headcount within the next six months. Whether the approach that worked for twenty people still works at forty is the next real test.

There's a lesson here that translates directly to small teams in Korea too. Testing a new feature on the tool your own team uses every day before putting it in front of the market — and showing it to early users even when it's rough — is essentially the entirety of what Labs actually did. It also suggests that cutting down approval steps matters more, at first, than adding headcount.

In the end, the real story here isn't the org chart — it's the sequence of events. Anthropic didn't plan out a product and then assign a team to it; it built something the model had just become capable of doing, watched to see where it got traction, and only then put people behind it. A tool that half the company's own engineers started using within five days was already proof that it would sell — and reading that signal required no market research and no approval process at all.

If headcount doubles over the next six months, the team will be erasing the very conditions that made it successful in the first place. Anthropic is now testing the shelf life of its own winning formula, and the result will show up in wherever the next product after Claude Code comes from.