If you're weighing whether to build an AI software factory in-house, this is the build log to read first. "The coding agent is the part you do not build." The project is the seven stations around it: isolated execution, snapshot templates, credential handling, git and PR orchestration, a server-owned run loop, live observability, and verification. Verification is the one everyone underestimates, since a green CI run tells you nothing about whether a user-facing change actually works. Two costs that never reach a budget line. Review attention grows faster than anything on an invoice. And an internal factory is a product with exactly one customer, so somebody owns it for as long as the company runs it, and that name usually gets decided by accident. The other useful reframe: running cost scales with how many tasks you queue, not how many engineers you employ. sandeep.dev/blog/what-a-soft…