Tech • AI • Robotics • Game
How Factory Is Building Autonomous Software With Claude
7/10TL;DR
Factory is building autonomous software agents for large enterprises, aiming to cut engineering drudgery, modernize legacy systems, and eventually run a near-continuous software factory with far less human oversight.
KEY POINTS
From string theory to enterprise software
Matan moved from a PhD focused on AdS/CFT, 10-dimensional black holes, and string theory into startup building in 2023 after becoming absorbed by program synthesis, now widely framed as code generation. The shift reflected a judgment that the hardest and most consequential problems would be found not in academia but inside large companies with sprawling, mixed-age codebases.
A mission centered on autonomy
Factory says its core mission has remained constant since its earliest pitch: bring autonomy to software engineering. The premise is that highly skilled developers spend too much time on repetitive, low-leverage work, especially in big enterprises where code can span 30 years, mix languages such as COBOL and Python, and require constant maintenance across old and new systems.
Why timing mattered
The company’s early years were described internally as a period of being ahead of market readiness. That lag became an advantage, giving time to learn hiring, product, engineering, sales, and operations before model capabilities caught up with the original vision of software agents that can work independently for extended periods.
The “dark factory” model
Factory is designing for a future in which software agents operate almost continuously, handling tasks for days or weeks with limited intervention. The goal is a system that can monitor deployed changes, detect deteriorating metrics such as a drop in completion rates, open a sandbox, test potential causes, propose a fix, deploy it, and keep monitoring performance.
Trust as the central product challenge
A key barrier is behavioral, not just technical. Enterprises must become comfortable delegating meaningful work to systems they do not watch line by line. Factory’s approach is to meet developers where they are, let organizations configure agents to fit internal processes, and earn trust by producing useful outputs before asking teams to increase autonomy.
Legacy banking as a proving ground
One of the clearest demonstrations came in banking, where legacy COBOL modernization projects often sit on backlogs because few engineers want to own them. In one case, an initial migration involved close collaboration between Factory engineers and bank engineers, while the next similar migration required 90% less human involvement because the system had learned the bank’s patterns and standards from the first pass.
Where frontier models matter most
The company argues that the biggest gains come from better planning and orchestration rather than raw code output alone. More capable models are being used for complex planning across many moving parts, reducing the need for constant review and making it more realistic to trust agent-generated plans without checking every step.
Opus and Fable reduced oversight burdens
Tools such as Opus and Fable are credited with taking over reliability-heavy work that previously consumed large amounts of human time. The practical effect is not just faster output but a change in management bandwidth: teams that once could ship only a few initiatives due to staffing limits may be able to launch far more, shifting the real question from capacity to strategic value.
Partnership and safety
In its work with Anthropic and Claude, Factory emphasizes reliability, clear division of responsibilities, and safety constraints that must be treated as immediate operational issues rather than distant concerns. The view is that these systems are already powerful enough to be both beneficial and intimidating, so deployment requires technical ambition alongside strong guardrails and careful communication.
Three bottlenecks to full autonomy
Factory sees progress depending on three layers: the model, the harness around the model, and the company workflows into which both are inserted. Model improvement over the past two years has been sharp enough that tasks once requiring both a model and substantial scaffolding can now be handled more directly, forcing companies to constantly rebuild their assumptions instead of letting tooling stagnate.
CONCLUSION
Factory is betting that advances in model reliability and enterprise integration will turn software engineering from a labor-constrained function into a far more autonomous system. If that happens, the main human role will shift away from repetitive implementation and toward deciding what is worth building in the first place.
Ask a question