At a glance
WHAT IT’S REALLY ABOUT
Factory’s “dark factory” vision: autonomous droids powered by Claude
- Matan Grinberg recounts moving from string theory PhD research to founding Factory after becoming obsessed with program synthesis/code generation and realizing the hardest problems live inside messy enterprise codebases.
- Factory’s mission is to bring autonomy to software engineering via “droids” that can operate for days or weeks, continuously monitor production impact, test fixes in sandboxes, and redeploy with minimal human intervention.
- A standout case study involves automating COBOL modernization at a bank, where after an initial learning-heavy engagement, subsequent migrations required about 90% less human effort.
- Frontier models like Claude Opus are highlighted as key to reliable long-horizon planning, enabling substantially less human oversight and unlocking dramatically higher shipping capacity (e.g., 300 changes vs. 3).
- Grinberg describes the remaining path to “dark factories” as a set of concentric blockers—model capability, the engineering harness around the model, and organizational workflows—requiring constant iteration and strong partnerships focused on reliability and safety.
IDEAS WORTH REMEMBERING
5 ideasFactory is building an “autonomous software factory,” not just a coding assistant.
Factory’s core bet is that agentic systems will increasingly handle end-to-end engineering work (planning, execution, monitoring) so humans can focus on higher-leverage decisions and domain context.
Being early can look like being wrong—but it can buy time to become the right founder.
Grinberg describes a learning period where the market and tooling weren’t ready, but the delay let him develop the company-building skills needed to scale once capabilities caught up.
The product is designed to change developer behavior through gradual delegation and measurable feedback.
Trust grows when agents run for days/weeks, produce meaningful output, and close the loop via deployment monitoring (detect regressions, test in sandboxes, propose fixes, redeploy).
Reusability and institutional learning can turn “dreaded” legacy work into compounding automation.
In a banking COBOL modernization, Factory’s system learned organizational patterns from an initial high-touch engagement; subsequent migrations required ~90% less human involvement.
Frontier models matter most for planning and orchestration because they reduce oversight cost.
Grinberg emphasizes that stronger models improve planning quality, making it feasible to reduce (or even skip) detailed human review except in rare corner cases.
WORDS WORTH SAVING
5 quotesThe mission that we have in the very first deck that we ever made is the same as our mission today, which is to bring autonomy to software engineering.
— Matan Grinberg
You've mentioned this concept of these droids working in a dark factory, lights off twenty-four hours a day, seven days a week.
— Kai Gill
And then after that, the next one, the amount of human involvement was like ninety percent less. That to them was like a magic moment.
— Matan Grinberg
Every company wants to ship 30 things, but they can only ship three in a quarter because of the resourcing that they have. Well, now you can ship 300 things.
— Matan Grinberg
We cannot settle. Like things are moving way too quickly. Like we need to always have like in one hand a sledgehammer to whatever we thought was like the best solution and like every month we need to rethink.
— Matan Grinberg
High quality AI-generated summary created from speaker-labeled transcript.
