A Juejin article this week confirms what Andrew Ng said earlier: 90% of AI Agent projects stumble at deployment — the problem isn't the model, it's the workflow. Specifically, three "cards" are missing: quality standards, feedback mechanisms, and stopping conditions.
What this is
The author proposes a framework called "Loop Engineering" (an engineering method for letting AI work autonomously in loops), using an Android login module as the case study — the module every engineer knows cold and whose requirements can be listed with eyes closed. The framework breaks down into three cards:
Card 1: What counts as qualified? Break "good" into verifiable clauses — "compiles successfully," "state is observable," "persists across token restarts." The moment any criterion is written as an adjective (e.g., "nicely written"), the AI will fake-pass on the first round.
Card 2: How do you give feedback when it fails? Feedback must be precise to the level of "which file, which line, what's wrong, how severe." A vague "try again" is equivalent to telling the AI to spin in place.
Card 3: When do you stop? Combine three triggers: all standards met, OR no new issues in two consecutive rounds, OR a hard cutoff at 5 rounds. Any one alone is insufficient.
Two more key design principles: judgment authority must go to an independent review Agent — never let AI self-audit (that's like letting students grade their own exams); the deliverable list must lock down core files but leave details flexible.
Industry view
The mainstream narrative: the AI coding race is about model capability, but at deployment, countless teams get stuck on "I can't even tell if what the AI wrote is actually working." This three-card framework makes the invisible role of "judgment authority" explicit.
We should also hear the dissent. Some engineering teams call this over-engineering — applying this process to simple script generation actually slows things down; an independent review Agent doubles costs, which small and mid-sized teams can't justify. The sharper critique: this method only works if you already have domain judgment — precisely the scarcest capability in the AI era. In other words, it rescues people who could already write a good login module; for newcomers, it raises the bar.
Impact on regular people
For enterprise IT: Before purchasing AI coding tools, align internally on what "qualified" means. Without verifiable acceptance criteria, even the strongest tool is just burning money.
For individual careers: A programmer's core competence is shifting from "knowing how to code" to "knowing how to define the problem and write acceptance criteria." People who can articulate what "good" means are worth more than those who just type fast.
For the consumer market: Consumer AI coding products will still market themselves on speed in the short term, but enterprise buyers will increasingly track two hard metrics: deployment rate and rework rate.