The useful clue in Catherine Jue's career is the list of things other people call boring. Bookkeeping. Payments infrastructure. Browsers. Each sits behind a more visible experience, humming away until something breaks. Jue has built companies and teams around those hidden layers, where a small improvement can travel through thousands of other products. Her latest company, Kernel, begins with a plain observation: an AI agent that needs to use the internet still needs somewhere to click.
The web was designed for eyes and hands. Its important actions live behind login screens, forms, dashboards and buttons. APIs expose only a fraction of them. An agent may be able to reason about a task, but it still needs a browser session that starts quickly, survives across steps, keeps credentials contained and lets a person intervene. Kernel turns those requirements into a developer platform.
Jue's route to this particular layer runs through two Y Combinator batches nine years apart, a financial technology career, a Japanese tea business and years of competitive Taekwon-Do. It is less a pivot-filled founder biography than a study in repeated form: notice the overlooked work, understand its details, then remove the friction without erasing the craft.
The first machine in the back office
In 2016, Jue was doing the books for her father's small business and found a world of manual categorization, reconciliation and month-end waiting. She had studied linguistics and computer science at UCLA, with interests in natural language processing and artificial intelligence. The mismatch was obvious: the technical world promised intelligent systems, while ordinary bookkeeping still asked people to move numbers by hand.
She and Ashwin Kumar co-founded Sway Finance and entered Y Combinator's Summer 2016 batch. Sway lived inside Slack and connected to services such as bank accounts, Stripe, Gusto and Expensify. A business could ask about sales, receive balances and generate real-time statements. Ten companies joined its early pilot. The ambition was familiar today, but the toolkit was not. Jue later recalled training their own models and wiring Node servers to Jupyter data pipelines in an era before broadly available language models and modern AI APIs.
Sway did not work out. That clean admission matters because Jue carried its lesson forward without turning it into mythology. She moved into engineering and product work at Nova Credit, founded Tekuno in 2019, and joined Cash App, where she became an engineering manager. The first company supplied a durable reference point: the idea may be directionally right while the available machinery is still wrong.
Nine years later, the bottleneck moved
The change became tangible inside Block. After GPT-4 arrived, Square CEO Alyssa Henry gave the company enterprise access and encouraged teams to use AI in their work. Jue's team built an internal question-answering product she described as “Glean, before Glean.” It sometimes hallucinated and the evaluation methods were immature, but it worked without a research lab. A long-held belief had crossed into practical software.
Her test had a playful beginning too. A friend had previously demonstrated DALL-E, and they asked it to draw cats in the ocean participating in a Japanese tea ceremony. Soon the same class of technology was retrieving company knowledge. The distance between curiosity and useful deployment was collapsing.
In February 2025, Jue left Cash App. She wrote that waiting even one more year to start a company could have meaningful consequences. With Rafael Garcia, the former co-founder and CTO of Clever, she began following the second-order effects of agents. Models were improving quickly. The tools around them were not.
Why the browser became the product
Browsers look like finished consumer software. Jue sees them as the universal runtime for work. Documents became Google Docs, accounting moved to QuickBooks Online, design moved to Figma and email moved to Gmail. The browser absorbed the desktop, but retained its document-era architecture. Now agents are being asked to operate inside that layered, irregular environment.
Kernel's first public technical move made the thesis concrete: the team put Chrome on a unikernel, a small, specialized machine image, so an isolated browser could boot in milliseconds. The open-source project reached the front page of Hacker News, gathered hundreds of stars and multiplied the company's small Discord community tenfold. Developers had confirmed the pain before the company had finished explaining the category.
The product expanded around production realities: persistent sessions, live views, recordings, secure credentials, human handoffs and support for tools such as Playwright. In October 2025, Kernel announced $22 million in Seed and Series A funding led by Accel, with Y Combinator, Cintrifuse Capital, Vercel Ventures, Refinery Ventures and SV Angel participating. Jue also named customers including Cash App and Rye.
Milliseconds change what gets built
Jue's writing returns often to speed. She argues that latency changes behavior: faster deployment makes developers ship more often, real-time streaming makes remote work possible, and instant financial settlement makes a transfer feel qualitatively different. Speed also reveals architecture. Dependencies and network calls cannot hide when every delay is visible.
For agents, this is not a cosmetic preference. A browser that takes tens of seconds to start may be acceptable for an overnight job. It fails inside a voice conversation or an interactive research tool. Kernel's early demonstration of browser sessions in roughly 20 milliseconds showed how an infrastructure metric could open a product category. The company has since explored WebRTC live views and GPU-accelerated browsers for graphics-heavy sites.
Her phrase is short, but the underlying standard is demanding. Making software feel simple often requires moving complexity out of the user's path and into the system. She saw that at Cash App, where extra steps were added only when necessary and messiness was handled behind the scenes. At Kernel, the same instinct appears in cold starts, authentication and observability.
A browser needs permission, not only performance
If agents can click, purchase, submit and schedule, speed is only half the problem. Jue's broader ambition is a web that can recognize legitimate agents. Her definition begins with consent: an agent delegated by an end user should be explicit, traceable and revocable. The website on the other side should be able to tell authorized activity from abuse.
Kernel has moved in that direction with managed authentication and work on identity and permissions. The position is practical. The web will not pause while every service agrees on a new interface. AI is useful precisely because it can bridge systems built at different times for different purposes. Browser infrastructure becomes the meeting point for old interfaces, new reasoning systems and the person whose authority makes an action legitimate.
The craft behind the cold start
The harder-to-copy part of Jue's profile may sit outside software. She is a Taekwon-Do world champion who won gold in patterns at the 2014 world championships in Rome, after two junior world titles listed on her professional profile. Patterns reward exact sequences repeated until technique holds under scrutiny. It is difficult not to see the rhyme with infrastructure, where reliability is practiced detail.
Tekuno reveals another tempo. Jue founded the San Francisco tea company in 2019 after working at Song Tea and joining a sourcing trip. Six years of Japanese study helped her build relationships with small producers in their language. She described each morning's tea as a sensory sequence: water boiling, leaves entering the pot, steam blooming, aroma and taste. Her collection of handmade teaware fills a bookcase wall, and she chooses a pot and cup to fit the mood.
Tea asks for slowness. Kernel asks browsers to start almost immediately. Both reward preparation so attentive that the final experience feels natural. The ritual is visible in one and hidden in the other.
Jue's personal writing adds small counterweights to the serious systems builder. Her site describes her, with dry humor, as “a very serious person.” It also points to rooftop gardening, Joan Didion, Annie Dillard and collections of visual fragments. In a Brooklyn rooftop garden, she documented the expense and uncertainty of filling a barren 500-square-foot space with grasses, asters, sedum and sumac. She admitted impatience, asked for help, and learned the plants' rhythms one season at a time.
Kernel is moving at the other end of that speed range. By July 2026, Jue wrote that the team had grown from two co-founders to 20 people in 16 months and had powered tens of millions of browser launches that year. The long view, however, remains ecological. She wants an internet where agents and humans can both participate, where delegation is legible, and where the open web's original values survive a new class of user.
That future depends on many unresolved questions. Jue appears comfortable there. She once described the excitement of technical work as not knowing whether an experiment will succeed. The browser, after decades as a familiar rectangle, has become uncertain again. For a founder drawn to hidden systems, that is a useful place to begin.