What a software factory is: four observable properties
Factory.ai’s definitional guide: standardized inputs/tooling, measurable output, replayability—and how agent-native factories change the throughput bound.
Source: Inside the software factory, and what it takes to build one
History nod (Bemer, Hitachi, Cusumano) then a crisp test: standardized inputs, same gateway for human and agent changes, measurable pipeline metrics, replayability from prompts/models/policies. Agent-native shift changes the actor inside the pipeline, not the pipeline idea—so incremental adoption is possible.
Throughput bound moves from engineer headcount to rate of intent definition, trace review, and release approval. Adoption sequence: scoped IDE delegation → automated PR review → multi-day missions. Build-vs-outsource argument: compounding context should stay yours.
Customer outcome snippets (Chainguard, Groq, Empower, Nav) illustrate velocity claims with the usual vendor caveat.
Why it matters for code review
Gives acceptance tests for whether your “factory” is real. If agent and human changes don’t share review gates and replayable evidence, you have acceleration without a production system.