Josh Albrecht and the Case for a Computer You Can Direct

The bargain was simple. Josh Albrecht wanted a new computer. His father said he would pay for one if Josh worked through a C++ book of roughly a thousand pages. The middle schooler did it. By the time he had a better machine, he had discovered a better pastime: writing games was more fun than playing them. In high school, he was paid to program them. A childhood deal had become a job, and it left him with a question that still shapes his work: how much more could people do if their computers met them halfway?

Albrecht is now the co-founder and chief technology officer of Imbue, a San Francisco AI company whose products include coding agents and tools for managing them. His ambition is broader than making code appear quickly on a screen. He wants people to be able to direct software, understand its behavior, and adapt it to their own purposes. The prospect is appealing even to someone who has never opened a terminal. For a researcher, an artist, or a small team, custom software has often meant waiting for someone else to build it. Albrecht wants that distance between an idea and a working tool to shrink.

A detour through the nature of thought

At the University of Pittsburgh, Albrecht first considered computational neuroscience. What held his attention was the larger question beneath it: how does thinking work? He moved into AI research with professor Rebecca Hwa, publishing work on machine translation and ways to understand language. He earned bachelor’s and master’s degrees in computer science. Those early papers belonged to a quieter era of machine learning, before conversational systems turned questions about intelligence into boardroom talk. His curiosity was already there.

He also wanted his work to meet the world. Research could answer questions, but he wondered how many people would use a tool he built or read a paper he wrote. Startups offered a different test. An early project involving anonymous BitTorrent browsing attracted users without becoming much of a business. Next came CloudFab, a marketplace matching people who wanted objects made with people who had 3D printers. Its work later included machine-learning estimates for injection molding, and the company was acquired. Even a factory quote, it turns out, can hide an interesting problem in prediction.

He later helped build financial software company Addepar as an early engineer and mentored Thiel Fellows. With Kanjun Qiu, he co-founded Sourceress, an AI recruiting company that joined Y Combinator’s summer 2017 batch. Recruiting looked like a matching problem at first: people, jobs, skills, preferences. Albrecht and Qiu learned that the human side was harder to compress into a model. A hiring manager’s idea of a good candidate might rest on credentials, experience, or an ability that a resume barely captures. The company built ways to represent those differences and handled sourcing and outreach. It also exposed the awkward incentives that can accumulate around hiring.

“Machine learning is not really about automation. I think it’s much more about augmenting people in the tasks they’re actually doing.”Josh Albrecht, in a 2021 interview

That sentence sounds like a statement of principle. It was also a conclusion drawn from operating a business. Sourceress worked, but the founders came to doubt that it could scale into the impact they wanted. They turned toward a wider research question: could they build AI systems that learn and reason well enough to be useful across many tasks, while staying responsive to people? The work grew into Generally Intelligent and, in 2023, took the name Imbue. Qiu became CEO; Albrecht remained CTO.

The large bet and the small interface

Imbue announced a $200 million Series B in 2023, with investors including NVIDIA and the Astera Institute. The money supported substantial model training and research on agents that could reason and code. Its planned computing cluster was on the order of 10,000 NVIDIA H100 GPUs. Those numbers made for a striking announcement. Albrecht’s account of the work, however, keeps returning to a much smaller unit: a person trying to get a computer to do something useful.

A model can produce an answer that looks plausible and still be wrong in ways a user cannot see. A coding agent can draft a feature and miss the test that would reveal a bug. For Albrecht, the design problem therefore includes the interface around the model: how a person reviews work, steers the next step, and decides when the result is dependable. At a 2025 San Francisco summit, he described Imbue’s approach as keeping engineers and domain experts involved in guiding and validating agent output. The hoped-for audience includes scientists and researchers who know their fields but need better ways to make custom software.

This is also where his interest in the personal computer becomes more than nostalgia. The old promise was that a machine on your desk could respond to your imagination. In practice, most software arrives as a set of choices made by somebody else. Albrecht sees agents as a possible way to let people build around their own needs. That requires the agent to remain answerable to its user. Openness, inspectable work, and the ability to move between tools matter because they determine who gets to make the choices later.

A product ships; a company takes stock

The principle had a difficult internal test in 2025. Imbue pushed to ship Sculptor, its coding-agent environment. The team worked late, added features, fixed bugs, and released it in October. People downloaded it; some returned week after week. Then came the ordinary work of a live product: user interviews, performance issues, regressions, alerts, and a backlog that did not care how pleased anyone was to have shipped. It was a real achievement, followed immediately by more work.

In January 2026, Albrecht wrote openly about what that stretch had done to the company’s habits. He and Qiu had been driving toward shipping and optimizing process, while the team’s sense of play and independence had thinned. His essay reads less like an announcement than a founder taking notes after an experiment. Imbue had said it wanted to increase human agency. Could its own staff choose and lead more of the work? Could smaller teams publish open projects that mattered to the mission, without waiting for one central roadmap to give them permission?

“Imbue is an experiment. It is a question: what happens if we give people agency and control and choice?”Josh Albrecht, 2026

The proposal was concrete: more team-led projects, clear ownership for the people driving them, and more tools built in the open. It also revealed something about Albrecht’s temperament. He is willing to put a company’s uncertainty in public, including his own part in it. Founders often describe a clean line between vision and execution. His account described the crooked line: make something, observe what happened, and adjust the way people work. The method resembles software debugging, with the unusual complication that the system contains your colleagues.

When one agent becomes a crowd

The next product offered a technical version of that same puzzle. In April 2026, Imbue released mngr, an open-source command-line tool for creating and managing coding agents on local machines or cloud computers. Albrecht discussed it with LangChain CEO Harrison Chase in a conversation recorded at Imbue headquarters. Their stated subject was how to usefully run as many as a thousand agents in parallel. The number catches the eye; the practical issues are more revealing. How do you know what each agent changed? How do you review the combined work? Who can see the state of the job, stop it, or move it elsewhere?

The appeal of parallel agents is easy to grasp. A hundred repetitive tasks could, in theory, happen at once. The burden of judgment remains. If every result requires guesswork to understand, the human has merely traded one queue for another. Tools such as mngr give Albrecht a way to explore that operational problem in public. The project is free and open source. It lets developers choose agents and computing environments instead of treating one provider’s workflow as the only possible future.

Albrecht’s other role adds a different vantage point. He is a general partner at Outset Capital, investing in founders while continuing to build at Imbue. His personal site lists interests that sound less like a standard investor biography than a shelf in a curious home: short speculative fiction, games, the mechanisms of thought, and the future of technology. He also helped create the Archive, a San Francisco co-living house. The through line is an interest in environments, physical and digital, that change what people can attempt together.

The old computer bargain still works as a miniature of the story. A parent offered a tool, with a demanding route to it. Josh discovered that the route could become the destination. He has since worked on translation research, factory estimates, recruiting, agent interfaces, and the organization of an AI lab. None of those fields gives an easy answer to how intelligence should be used. His recurring answer is to give people more say in the systems around them. The next machine may write code on command. The more interesting question is whether its owner can still call the tune.