Factory’s Matan Grinberg on building software agents people can trust

Matan Grinberg thinks technology can give people time to rethink existing work and find a new angle. Models such as Opus and Fable, he says, can take on work that previously occupied much of that time.

Anthropic’s Kai Gill asks Grinberg about his move from studying string theory and ten-dimensional black holes to founding Factory in 2023. Grinberg says he grew up in the Bay Area but initially wanted nothing to do with startups, so he spent about 12 years in theoretical physics.

A seminar on program synthesis, now commonly called code generation, captivated him. By 2023 he felt the most interesting problems required industry: large organizations had codebases combining decades-old systems, newer code, COBOL and Python. Factory’s mission from its first pitch, he says, was to bring autonomy to software engineering and free engineers from lower-leverage work.

Gill asks what Factory learned during the years when its vision of agents was ahead of the models. Grinberg says the time allowed him to grow into the job. It was his first job, and he knew little about hiring, engineering, product, sales or marketing; had the market been ready immediately, he thinks he might have missed the opportunity.

Gill asks about Factory’s vision of Droids, its coding agents, working in a “dark factory” around the clock. Grinberg says that as people delegate tasks taking days or weeks and see meaningful output, they begin to change their habits and build trust in work they do not watch continuously.

His example is a software system that detects a drop in completion rates after a deployment, creates an isolated test environment to investigate, and shows a person a candidate fix before deploying and monitoring again. Developers know their own businesses, he says, so Factory should let them configure the system to fit their organizations.

Grinberg describes a bank modernization involving COBOL, a language used in many older business systems. A Factory engineer worked alongside bank engineers on the first migration while the agent extracted lessons about how that bank worked. For the next migration, he says, human involvement fell by 90%; the bank’s engineers were delighted to spend less time on work they had dreaded.

Asked where agents need the strongest models, Grinberg compares the problem to a company’s CTO: having that person write every line of code would waste their broader context and problem-solving experience. He uses Opus when planning a complicated job with many moving parts.

For Factory, he says, the main benefit of Opus is less human oversight. A bad plan for a long task may only become apparent much later, but he has become comfortable enough to accept plans without always reading them closely. He says he now revisits a plan only in rare, unusual cases.

Some company work has to meet a reliability threshold and used to consume substantial time. Grinberg says Opus and Fable can take on much of it. If a company can ship far more things with the same resources, he asks, will those things actually make customers happier? In his view, people must decide what the company does best and how to deliver it better.

Grinberg finds that shift exciting because people can reconsider assumptions about what they can build. He sees more time to approach a problem from a new angle when models take over tasks that previously filled the working day.

On Factory’s partnership with Anthropic, Grinberg emphasizes reliability, shared priorities and agreement over who should work on which problems. He says the partners want to expand human potential while recognizing safety constraints and taking care in how they explain powerful tools.

Those tools can be frightening because they are powerful, Grinberg says, but he also wants to highlight their benefits and how Factory’s Droid and Claude can improve together. He regards safety as a present problem across AI engineering, not just a concern for a future system.

Asked what still blocks the dark factory, Grinberg describes three concentric circles: the model, the surrounding software that makes it useful, and the company’s workflows. All three are changing quickly. He says a model today can do things that two years earlier required a model plus additional supporting software.

As models become more capable, they can absorb functions previously performed by that surrounding software. Grinberg argues that teams should repeatedly question whether their setup is still optimal, rather than assume a solution that worked last month remains best. A close model-provider partnership can help them understand which gaps to fill in the meantime.

For prospective founders, Grinberg says he usually tells people not to start a company if they need his affirmation. His own preference is to become absorbed in a problem: find something that keeps you thinking at night and draws out your curiosity. He believes today’s tools make even some unusually ambitious ideas more achievable than they once were.