The Browser Company ships fast because work has nowhere to get stuck

You have probably watched a small team ship in a week what your organisation debates for a quarter. The usual explanation is size or talent density. The Browser Company of New York, the startup behind the Dia browser and its earlier iteration Arc, has 65 people and an AI-native toolset, but the bigger factor is a way of collaborating it designed on purpose to keep work from getting stuck. Researchers from Atlassian's Teamwork Lab, which Dr Molly Sands leads, interviewed 10 Browser Company employees in June 2026 to find out how that system works. Atlassian now owns The Browser Company, so read this as an inside look from the parent company.

Atlassian's State of Teams research calls the underlying problem the fragmentation tax: the reviews, approvals and handoffs that make up roughly 80 per cent of knowledge work. Most AI gains die in those gaps, and flooding the gates with more AI output only chokes the system further. The Browser Company is what happens when a company removes that tax on purpose. Three practices explain the pace.

Hiring for judgement

The Browser Company recruits what it calls "un-blockable" people, flexible generalists comfortable working outside their own specialty, and it judges every candidate against a bar built for the AI era: Independence, Instinct and Impact. Independence asks whether the person can go two weeks without needing to be pulled out of a rabbit hole. Instinct asks whether they can find the highest-leverage thing to work on without being told, and the company says this is the hardest quality to hire for and the one it grows rather than screens for. Impact asks whether they ship the right things.

AI then makes these "lane-swervers" faster still. The cross-functional range that used to take years to build is now within reach of anyone willing to try, and AI tools give people the confidence to attempt work they would once have handed to a specialist. Co-founder and CTO Hursh Agrawal is blunt about why independence is non-negotiable. "We just can't support anybody who's not independent, mostly because we as managers are finding it's so much more productive to tell AI what to do than to tell people what to do," he says. "If there's ever a question, it turns out it's way better for the manager to just be telling AI and doing it themselves. So we really demand that." AI simplifies execution, and direction still has to come from people, which makes judgement the real speed multiplier.

Shrinking the unit of ownership

Most organisations lose time in the spaces between teams: review queues, approvals, context gaps and ownership boundaries. The Browser Company keeps those spaces small in three ways. Everyone shares context and feedback openly through constant dogfooding and direct, in-the-moment critique, so the whole team works from the same information. Tiny pods, often just one to three people, own work end to end, and the company pushes each decision to the smallest possible group with one clearly named person accountable for the outcome. Leadership sets one clear North Star each season (this summer it was proving the business model, distilled into blunt mandates like "prove $5/day of value") and then hands pods full autonomy on how to hit it.

The result is that AI-generated output rarely piles up waiting for another team to absorb it. "Really small working groups of people who are really individually empowered is the real secret sauce," says Ben Cunningham, a design engineer at the company. A 65-person pod model will not lift cleanly into a much larger organisation, and the Teamwork Lab researchers say as much. The lesson that travels is diagnostic: if your AI gains are disappearing, look at your handoffs before you blame your tools.

Building to settle debates

When an idea is contested or there are competing directions, the team builds something testable instead of arguing in the abstract. That works because AI-aided experimentation is cheap, internal feedback loops are fast and the people building the product use it heavily. In practice the team resolves in days decisions that elsewhere turn into weeks of alignment.

One of Dia's marquee features shows how fast that can run. A non-engineer spent a few hours prototyping an idea with AI and shared it internally. By Thursday, leadership had named it a bet. The following Monday a small team kicked off, and about a month later Dia's full agent experience shipped with a new tech stack, a new model and a new security architecture. Nothing sat in a queue along the way.

The same instinct makes reversing cheap. The company treats features, processes and even its planning cadence as prototypes, shipped to learn and killed without ceremony. It moved from six-week cycles to quarter-long seasons when the shorter rhythm stopped fitting how people actually worked, and it dialled back Dia's report-generation feature after both users and the team pushed back. When a bet does not land, leaders own it in the open. One employee told the researchers there is "no finger-pointing I've ever seen at the company", and "strong opinions loosely held" is how the team stays aimed at the right target while AI keeps changing what is possible.

What this means for L&D

The Browser Company's system works because it aligns three things: the people it hires, the way it organises work and the norms it uses when decisions stall. Most organisations try to fix velocity by changing one of those in isolation, with a new project-management tool, a flatter org chart or a workshop on decision-making, and then wonder why nothing changes. The Teamwork Lab's conclusion is that the bottleneck is shifting from production to decision-making, and the teams that get the most from AI are redesigning hiring, team structure and decision-making so faster output does not get trapped inside slower systems.

For learning and development leaders, the most portable lesson is the economics of prototyping. AI has made it cheaper to build a rough version of the real thing than to align on a description of it, so teach teams to answer contested questions with a prototype and give them a named owner who can make the call. Then pick one place where work piles up between people, give a small group end-to-end ownership of that problem for a fixed period and measure what happens to cycle time. Hiring for judgement takes longer to change, but the independence test is a useful screen for any role that now works alongside AI: can this person go two weeks without someone pulling them out of a rabbit hole?

Sources:

• Atlassian, October 2026