Source: raw/How_Factory_is_building_autonomous_software_with_Claude.md — official Anthropic YouTube channel (@claude), interview with Matan, founder of Factory, youtube.com/watch?v=N6FkmdL0Lg0 (fetched 2026-10-30). The interviewer is not named in the transcript.
Factory sells “Droids”, coding agents for large enterprises, with the mission (unchanged since its first pitch deck) “to bring autonomy to software engineering.” In this short interview Matan explains the “dark factory” goal (agents working lights-off, 24/7), the bank COBOL migration that cut human involvement by 90% on the second run, why he routes planning to Opus, and his three-circle model of what limits autonomy today: the model, the harness, and the company’s workflows. The figures are Matan’s own and are not broken down further.
Key Takeaways
- The target is a “dark factory”. Droids working “lights off 24 hours a day, 7 days a week”. Matan says users get there through behaviour change: once they delegate tasks “that can go for days or weeks on end”, they get used to not touching something for a couple of days, and they build trust when the output turns out to be meaningful.
- The monitoring loop he describes. The factory is “always monitoring”: a deploy lowers the completion rate through some step, so a Droid spins up a sandbox to find out why, finds a candidate fix, pings a human with it, deploys it, and monitors again.
- Configurability over opinion. “Respect their intelligence as developers. They know their business more than we do.” The product should let customers “morph it to the silhouette of their organization.”
- The hero case is a COBOL migration at a bank. Every bank has a backlog of COBOL modernizations that no engineer wants. On the first one, a Factory engineer sat with the bank’s engineers and the Droid “was extracting lessons” as it worked (“this is the way this bank does things”). On the next migration, “the amount of human involvement was 90% less.” Matan describes “tears of joy” from the bank’s engineers.
- Route planning to the strongest model. Matan compares it to not having the CTO write every line of code: the person with the most context designs the general solution. “If I’m doing planning or something pretty complicated where I need to orchestrate a lot of moving parts, I’ll use Opus.”
- Less oversight is the main thing Opus gave them. Planning mistakes are expensive because “you’ll only find out if you went awry, billions of tokens down the line.” With Opus he trusts plans “to the point where you literally might not even look at them”; going back to check a plan now happens only in “a very crazy corner case.”
- Three concentric circles limit autonomy. At the core is the model, then the harness, then “the company and the workflows of the company as a whole.” All three are growing quickly. “A model today could do what a model plus a harness could do 2 years ago.”
- Models absorb the harness, so keep a sledgehammer handy. “If you are lazy, and you don’t change your harness, it’s going to be superseded.” His rule: “We need to always have in one hand a sledgehammer to whatever we thought was the best solution. Every month, we need to rethink.”
- Capacity changes the question humans ask. “Every company wants to ship 30 things, but they can only ship three and a quarter… now you can ship 300 things.” The question becomes whether those 300 things make customers happier, and what the company’s core competency really is.
- Being early cost two years. Factory started in April 2023, when “the models weren’t really there yet.” Matan calls 2023 to 2025 a period in which he learned hiring, product, sales and marketing for the first time; had the timing been right then, “I was the wrong founder.”
Try It
- Make a second run cheaper than the first. On a repeated migration or refactor type, have the agent write down “how this codebase does things” as it works (a lessons file or skill). Then measure human time on the second run against the first, the way Factory measured its 90% figure.
- Split planning from execution by model. Plan with the strongest model you have and execute with cheaper ones. Note how often you go back to correct the plan; that rate is your measure of how much oversight the planner saves.
- Add a monthly harness review. List the scaffolding you built to make up for model weaknesses (retries, extra checking, prompt workarounds), and with each model release test whether the model still needs it.
- Wire one post-deploy loop. Pick one metric that a deploy can break, such as a funnel step’s completion rate, and give an agent the job of investigating any drop in a sandbox and proposing a fix for a human to approve.
Related
- The Software Factory — Warp’s Closed Loop and Ramp’s Five Bottlenecks — the same “software factory” framing from Warp and Ramp
- Maintain the Harness, Don’t Pile On Tools — the opposing view that the harness is the durable layer to maintain
- How Anthropic Runs Large-Scale Code Migrations with Claude Code — Anthropic’s own migration method
- Block — Orchestrating Claude Fable Across Thousands of Pull Requests — another enterprise case of a strong model planning and cheaper models executing
- Patrick Collison on Claude Code at Stripe — Stripe’s version of long plans dispatched to many devboxes
- The Capability Curve — Anthropic’s view of models outgrowing scaffolding
Open Questions
- Matan’s surname and title are not given in the transcript. Factory’s CEO is widely reported as Matan Grinberg, but that is unconfirmed here.
- The bank, its migration’s size and how “human involvement” was measured are not stated. The 90% figure is one comparison between two migrations.
- Which Opus version Factory uses for planning, and which models run the execution steps, is not said. “Opus and Fable” are named only together.
- How the post-deploy monitoring loop works in the product today (shipped or aspirational) is unclear; Matan describes it as what the software factory should do.