André Staltz says to make agents fail closed and automate slowly

Staltz posted a numbered summary of a talk he says he gave recently. We read the post in full. We did not find a recording or slides, so what follows is his own summary of his talk. He defines a software factory as a system of agents that build software with humans occasionally in the loop, and says the engineering craft moves into agent instructions and infrastructure. His good reasons to build one are less tedious work, less slop, and more trust in what agents build. His poor reasons include automating everything as soon as possible and wanting to stop looking at code.

On review he makes three claims. Work out what you check when you review a pull request, then teach an agent to check those things and report on them. He says he now trusts bot reviews more than his own, because they are systematic and can be audited later, while his own review is ad hoc. And he says people should write and read English, with agents following a set reporting process when they describe pull requests, because authoring in instructions while reviewing in code makes review the bottleneck.

Two lessons close the list. Narrow each agent with rules and linting and make it fail closed: an agent that lacks a correct environment should fail, and the person should fix the environment. The second: "Premature automation is the root of all evil." He advises scaling up little by little and monitoring.

Staltz's pinned post announces a release of Fabrica, a tool for running a software factory, so this is advice from someone with a product in the category. The post gives no measurements. It showed 184 likes, 292 bookmarks and 14.2K views when we read it on the evening of November 11, about 27 hours after posting.

Why it matters: Fail closed is a rule a team can write into a harness this week: a broken test environment stops the run, and the agent is told not to work around it. The review claim needs more caution. Reading an agent's English report in place of the diff works only if the report is accurate, and the post does not say how he checks that.