Factory's Matan Grinberg: The Coming ‘Dark Factory’ Where Software Builds Itself

Summary

Factory started in April 2023, long before enterprises were ready for autonomous software agents, and Grinberg says the company spent its first two years in a “journey in the desert.” He recounts an unusually drastic reset: when the product was not creating delighted developers, Factory proactively refunded customers rather than keep revenue it hadn’t earned. The rebuilt product emphasizes model independence, keeping customer artifacts in the customer’s codebase, and a router that can assign generation, testing, and review to different models based on cost, reliability, and policy. Grinberg also argues that harness design matters more than model-harness co-design, that open-weight models are already good enough to capture most tokens, and that software development is moving toward asynchronous execution in a “dark factory” where humans increasingly specify goals and validation criteria rather than write every line synchronously.

Notes

• Factory was founded in April 2023, when enterprises were still barely ready for Copilot-style workflows, much less autonomous coding agents.
• He describes the first two years as a “journey in the desert,” with the market not yet ready for the product.
• Enterprise buyers care about model independence and modularity, not just raw model quality.
• Factory positions itself as a layer that preserves customer-owned work products and avoids lock-in.
• Automations, skills registry artifacts, and other outputs stay in the customer’s codebase.
• The company refunded customers when it concluded the product was still not good enough.
• Revenue had reached just under $2 million before the refunds.
• He reframes customer obsession as an input metric; the output has to be customers becoming obsessed with the product.
• The team stayed together because hiring was ruthless around mission obsession and feedback loops were clear.
• Factory’s droid CLI launched on September 26, 2025.
• The droid CLI was meant to meet developers where they actually work.
• Model-agnostic performance is a core product claim for the CLI.
• Developer behavior changed as much as model capability: engineers became more willing to try agents and tolerate guardrails.
• Benchmarks have a short half-life because they get gamed or saturated quickly.
• Terminal Bench favored terminal-native, general software-development tasks over narrow file edits.
• Harness performance depends on token caching, compaction quality, needle-in-the-haystack retention, tool use, and environment validation.
• Cash tokens are said to be about one-tenth as expensive.
• He says a harness that supports many models beats one co-designed with a single model.
• He argues harnesses should avoid overfitting to one model’s quirks.
• Factory’s router can assign different models to generation, testing, and review.
• Routing instructions can be provided in natural language.
• Internally, Factory says about half of its tokens now go to open models.
• He expects open models to be judged against the previous frontier generation, not the newest one.
• Open models are already shifting from marginal to meaningful usage at customers.
• He expects open models to dominate token share over time, while frontier models keep the highest-leverage decisions.
• Low-leverage work like documentation should be automated so engineers can spend more time on high-leverage reasoning.
• Factory thinks usage-based pricing fits today, while outcome-based pricing may arrive in the 2030s.
• Seat-based pricing does not fit Factory’s current product shape.
• He says model-independent routing creates something like a marketplace between engineers and model providers.
• Future pricing could become outcome-based bidding with validation loops.
• He thinks task subdivision and bidding mechanics remain an unsolved engineering problem.
• Modern enterprise software development is still highly inefficient and largely undocumented.
• He argues AI can codify tribal knowledge, approvals, and checklists into repeatable software-factory assembly lines.
• The next allocation question becomes dollars to headcount versus dollars to tokens.
• He wants business decisions to become more mathematical, not vibe-based.
• Sales teams likely need more human headcount than extra tokens, while engineering can get leverage from more tokens.
• He says many companies wrongly try to build software development agents despite that not being core to their business.
• Company-wide AI hackathons are an effective adoption pattern.
• Adoption works best when experimentation comes from within the org and failure is tolerated.
• He predicts 90% of tokens will be asynchronous within 12 to 24 months.
• The “dark factory” metaphor is central to his long-term software vision.
• He expects short-term turbulence as AI drives painful resource reallocation.
• Long term, he thinks software can absorb many more unsolved problems, including government systems and pharmaceutical research.
• He says the industry should reallocate top engineering talent quickly, ideally within six months or a year.