What they did
Eduardo Rosado describes the software factory his company Abant has run on its own product from July to November 2026. Agents pick up GitHub issues, work in isolated git worktrees with no person watching, and open pull requests that a different agent reviews. Each issue lists the files the change may touch and acceptance criteria that end with the repo's validation command, and a plan cannot be published as issues without both. An orchestrator compares the files a PR changed against the issue's list and holds the merge for a person if anything falls outside it. A session that finds the spec ambiguous prints one concrete question and exits; the task waits until someone answers. Per-file size and coupling budgets fail validation when exceeded. The agent products and models are not named.
How it went
Every session is told to run full validation before opening a PR, yet roughly one in seven PRs (about 14 percent) had at least one failed CI run. Rosado treats that as an upper bound, since some failures were flaky tests or differences between the agent's machine and CI, and the causes were not classified. His conclusion is that instructions alone did not keep changes green and the checks caught what slipped through. Nothing in the factory catches a change that is bigger than it needed to be.
Worth knowing
The useful split is between rules a machine can check (allowed files, acceptance commands), rules you can only make cheap to follow (stopping to ask), and rules left to the reviewer (simplicity). File-level scope checks still let an unneeded edit through inside an allowed file. The post is in Spanish, and the mix of models over four months means it shows no per-model trend.