Before there were Droids, enterprise deployments or a $5 billion valuation, there was coffee at noon. Eno Reyes and Matan Grinberg had spent their Princeton years in a peculiar near orbit: the same university, a thicket of mutual friends, even opposing eating clubs, yet no proper conversation. They finally connected in San Francisco at a LangChain hackathon in 2023. The coffee came the next day. Grinberg later described what followed as a conversation that never really stopped.
The speed of the pairing was slightly comic; the subject was not. Large language models could already produce code, but production software is less like writing a clever sentence and more like renovating a hotel without asking the guests to leave. Repositories contain years of decisions. Tests carry institutional memory. Dependencies sulk. Security teams ask sensible questions at inconvenient moments. The pair started Factory to make AI useful in that untidy world.
Reyes brought an engineer's view of the mess and an operator's suspicion of magic tricks. At Microsoft, where he worked in Azure Networking, and later at Hugging Face, where he helped enterprise customers work with language models, he watched capable teams lose time to internal tools, scattered knowledge and technical debt. His conclusion was blunt: typing code was not the hard part. Finding out what should be changed, why it should be changed and how anyone could know the change was safe was the harder craft.
“The hardest part of building software isn't writing out lines of code - it's navigating the maze of internal tools, dependencies, APIs, tech debt, and fragmented knowledge that slows teams down.”Eno Reyes
A family loop closes in San Francisco
Reyes grew up in Atlanta, but his route to San Francisco had begun a generation earlier. His paternal grandparents came from Puebla, Mexico, in the late 1960s. They opened a Mexican restaurant in Los Altos in the 1970s, then moved it to Haight and Cole in the 1980s before the family left for Georgia. When Reyes later moved into the Mission District for his job at Hugging Face, he found himself a block from where his grandparents had once settled.
He has spoken about their drive to build a successful life for his father and about the same mentality passing through his parents to him. The detail lends his return to the Bay Area a satisfying shape. San Francisco was both a new technical frontier and old family territory. Startup mythology normally adores the garage. Here, a restaurant and a neighborhood do more useful work.
At Princeton, Reyes studied computer science and wrote a thesis involving deep learning. He also started early. In 2017, as a freshman, he co-founded 23cubed, initially a tidy corporate wrapper around freelance website work for friends, family and whoever else needed it. Clients soon wanted search optimization, advertising and reputation management too. The side project grew into a broader digital agency.
Years before autonomous agents became his occupation, Reyes was already thinking about repeatable work. His student-era advice was practical: document any process repeated more than twice, research internal tools carefully and resist scaling either too fast or too slowly. Factory would pursue a much larger market, but it kept the same fascination with the hidden machinery behind visible output.
- 23cubed begins in a Princeton dorm room as a web-development agency.
- Microsoft brings Reyes into software engineering for Azure Networking.
- Hugging Face gives him a close view of enterprise demand for language models and code generation.
- Factory forms after the San Francisco hackathon reunion with Grinberg.
- $5 billion becomes Factory's reported valuation after a September financing.
The droid had to grow up
Factory's original idea was to divide software work into specialized autonomous systems: one might plan, another review, another test or document. The company called them Droids. The founders liked the word because “agent” had become associated with unreliable demonstrations. There was a small legal joke hiding in the origin story. They first incorporated as the San Francisco Droid Company, then changed the name after advice about the vigilance of Lucasfilm's trademark lawyers. Thus Factory: industrial enough to imply repetition, broad enough to avoid a lightsaber duel.
The early product broke work into recognizable stages because those stages gave an agent boundaries. In Reyes's language, they made the problem more like a game, with a state, available actions and feedback. Different tasks demanded different “cognitive architectures.” Code review and debugging might both involve software, but the route to a reliable answer was not identical. He resisted the idea that one reasoning recipe would rule them all.
context
work
changes
signals
learning
That remains the heart of his argument. A model can generate an answer. An agent operating for eight or ten hours must manage limited context, decide what environmental information matters, use tools correctly, interpret noisy test results and recover when a plan fails. None of those pieces supplies a cinematic secret. Together, they form what the industry calls a harness. “The sum of all of these things requires attention to detail,” Reyes has said. “There really is no individual secret.”
This is also why Factory is model-agnostic. Reyes does not want an enterprise to reorganize its engineering culture around the preferences of one model provider, operating system or editor. Models can be swapped; the surrounding state, verification, learning loops and workflows are where he expects durable value to accumulate. His version of technical sovereignty is not a flag planted on a data center. It is the ability to choose the intelligence, keep the organizational learning and control where the system runs.
The proof is in the old code
The least fashionable code may be the sharpest test. In April 2026, Reyes co-authored the introduction of Legacy-Bench, a collection of hundreds of tasks spanning six older language families and enterprise domains. The premise was sober: systems that settle financial transactions, route calls and adjudicate insurance claims still depend on COBOL, Fortran and older Java, while popular coding benchmarks concentrate on modern Python and JavaScript.
A benchmark cannot reproduce every argument in a meeting or every undocumented dependency in a bank, but it can force the agent toward the places where maintenance is consequential and expertise is thinning. It also reveals Reyes's taste in problems. He is drawn to large questions about intelligence and consciousness, yet his company keeps arriving at unromantic tests: Does the patch pass? Was a secret exposed? Can an old system be changed without breaking the next payroll run?
Factory's commercial acceleration has been severe. It announced a $5 million seed round in November 2023. By April 2026 it had raised a $150 million Series C at a $1.5 billion valuation. Five months later, a further $200 million financing valued the company at $5 billion and took disclosed funding above $400 million. Factory said its systems were used by hundreds of thousands of developers, with customers including Nvidia, Adobe, Palo Alto Networks, T-Mobile, Blackstone and Royal Bank of Canada.
The platform expanded with the ambition. Customers could run it in the cloud, in their own virtual private clouds, on premises or in air-gapped environments. A router chose models at the task level. An effectiveness product attempted to measure the value of the work. The vocabulary had shifted from individual coding agents to self-improving software factories. The recurring promise was control: over models, deployment, security and what the system learned.
“The companies that win will own the verification, state, learning loops, and workflows around the model.”Eno Reyes
A taste for very large questions
Reyes's personal website reads less like a polished founder biography than a cabinet of questions. Under intelligence and consciousness, he wonders whether learning systems converge on the same high-dimensional shape. Under art, he considers the cultural experiences minds need to share, beyond information. Under aliens, he entertains dark forests, invisible mass-energy and the sociology of institutional secrecy. The section on software asks what people might make if the number of creators grew from tens of millions to billions.
The range could look eccentric if the questions did not share a family resemblance. Each asks how intelligence becomes collective, how knowledge moves and what happens when a limiting resource becomes abundant. Even his admiration for Jeff Dean comes with a social qualifier: the research matters, but so does the report that Dean is a nice person. Reyes has emphasized collaboration in a field that often prefers its geniuses solitary and conveniently lit.
He was named to the 2025 Forbes 30 Under 30 list in artificial intelligence, but his public posture is less boy-wonder than systems mechanic. He talks about cost-quality tradeoffs, model provenance and liability. He argues that open models should be judged technically rather than by national shorthand. He expects most routine workflows to move toward open systems while a smaller share of difficult tasks captures a larger portion of economic value. It is a view of AI as a market of components, not a throne waiting for one model.
The aspiration, however, remains expansive. Factory was founded around a world in which software could build and improve itself. Reyes frames that abundance as a way to widen creative capacity rather than simply reduce head count. His childhood, his family's restaurant story, his student agency and his years inside large technical organizations all point toward the same preoccupation: people have ideas, institutions have friction, and useful systems turn one into the other.
Coffee at noon is a charming origin because it is so ordinary. The harder work came afterward: thousands of decisions about tools, context, evaluations, security and customers. Reyes's wager is that autonomy will arrive through that accumulation, less like a lightning strike than a factory line being tuned until the output holds. The future of software, in his telling, is not merely a machine that can write. It is a machine that can understand the assignment, live with the consequences and come back with the tests passing.