The factory model is Addy Osmani’s mental model for agentic software development: rather than writing code directly, a developer builds and runs a factory — a fleet of AI coding agents working in parallel, each with its own task, toolbelt (repositories, test runners, deployment scripts, documentation), context (specs, architecture decisions, prior constraints), and feedback loop. Instead of hand-holding one agent through one task, a developer spins up several agents at once — one on a backend refactor, another on a feature, another on integration tests, another on documentation — and reviews outputs, gives feedback, refines specs, and redeploys.
Usage
Osmani extends the factory analogy on several points he treats as holding directly: a factory has quality control, has process documentation, has inputs that must be precisely specified or the output comes out wrong, and stalls when its environment is unreliable. He reports that inside teams that have adopted the model aggressively, a substantial portion of merged pull requests now originate from agents running autonomously in cloud environments, presenting this as production reality rather than theory. He connects the model to public remarks from Cursor, quoting it as saying “the developer’s job is becoming building the system that builds the software, the factory, not just the product” and “reviewing ideas is a lot more fun than reviewing code.”
The “software factory” framing is also used as a product’s own self-description, independently of Osmani’s account. LangChain describes Open SWE as an open-source software factory and states its purpose as turning engineering work into a repeatable system, with a loop it draws explicitly: issues, conversations, pull requests and schedules feed planning and investigation, which feeds implementation in an isolated sandbox, which feeds validation and delivery of a pull request, which feeds review, CI and feedback — and follow-up work returns to planning. That is a narrower and more mechanical statement than Osmani’s mental model: it names a single agent system’s own control flow rather than a way for a developer to think about directing many agents.
When It Applies
Osmani presents the factory model as the appropriate approach once a developer is directing autonomous agents that can run largely unsupervised for extended periods, rather than a single synchronous agent handled step by step. It assumes a developer can write and maintain a specification precise enough to survive being run against by many agents at once, plus a comprehensive, test-first suite to check the results — without those, he argues, ambiguous requirements and poor architectural decisions propagate across the whole fleet rather than staying contained to one implementation. He presents it as his own mental model for this shift, while reporting it as already production reality inside some engineering organizations.
Related Terms
The Code Agent Orchestra - what makes multi-agent coding work, Conductor and Orchestrator Modes, Red/Green TDD