Wink Pings

AI Software Factory: Four Principles, One Side Effect

Mihail Eric laid out four principles for a software factory. In the comments, one reader spotted the side effect: when the number of human interventions becomes a metric, the easiest way to meet it is for no one to intervene.

Mihail Eric recently laid out the full set of components of what he calls a “software factory” on X. At its core, the idea is to have AI agents take on the coding tasks that development teams used to perform, with all tools, environments, and processes standardized like an assembly line.

He offered four core principles.

Agent readiness is the first gate.
An agent needs access to the tools required for the task, an isolated runtime environment, and a code repository it can start and verify on its own. If an agent cannot bring up a development environment without help, the factory cannot be built yet.

All changes go through the same gate.
Continuous integration (CI), static analysis, tests, and AI reviewers—whether a change comes from a human or an agent, it goes through the same gate. An AI judge scores each full run against the specification. The best benchmark is your own history: replay past PRs (pull requests) and compare different models and configurations in terms of cost and quality.

Every failure makes the system better.
When a defect slips through, record the reason and turn it into documentation, a skill, a lint rule, or a tool so that the defect rate decreases over time. An agent going off track is a signal that the factory needs improvement.

The human role shifts from worker to manager.
Plan synchronously, execute asynchronously, and conduct human reviews graded by risk. Humans still guard the gates for security, data, and payments. Then measure PR cost, cycle time, and the number of human interventions per PR—not lines of code or token count.

image

So far, the principles seem reasonable. But two replies in the thread are worth noting.

User Mykyta Pavlenko pointed out a metric trap: “Once ‘number of human interventions per PR’ becomes a performance target, the cheapest way to meet it is for no one to look at all.” This is the classic Goodhart’s law scenario: when a metric becomes a target, it ceases to be a good metric. In Mihail’s system, human review is aimed at security and payments, but if teams start staring at this number, review will become a formality, and agents will quickly learn to exploit it.

User olivier described a fascinating scenario: they have an AI playing World of Warcraft, and this AI constantly rewrites its own code. Once, the agent bound a new action to the key it normally uses for walking, so the character just kept holding down the forward key and stayed in place. Their approach was straightforward: no fancy tricks—if the game performs worse after a change, they roll back the change. This happens often.

olivier’s anecdote happens to validate Mihail’s third principle: an agent going off track is an improvement signal. But it also shows that, in the “software factory” framework, the cheapest way to “make the metrics look good” is often not to make the system better but to make the metrics meaningless.

发布时间: 2026-11-21 07:50