A stop condition is a piece of specification work, and it is not the kind that delegates well. In Jiao's account the check that governs an autonomous loop does not read source code at all; it opens the running software, uses it the way a person would and reports what actually happened [12]. His example is an image tag carrying perfect accessible alt text over a dead source URL, which passes every static check and fails the only question that matters, whether a person can see the image [11]. Somebody has to decide, in advance, which of those two signals the loop is allowed to believe.
The failure mode he names is drift. An open loop with no meaningful check can report done even when the code does not work, because nothing in the system is permitted to disagree with it [10]. Jiao adds two properties that read more like procurement requirements than engineering taste: the check has to resist gaming, since an agent will satisfy a loose check rather than the underlying goal, and it has to preserve what already works, so this morning's fix does not reintroduce last week's bug [12]. Both are constraints on the people writing the check rather than on the model doing the work.
Andrew Ng's three nested loops give the timing problem its shape, with agentic coding measured in minutes, developer feedback in hours and external feedback over days or weeks [5]. Put an inner turn at ten minutes and an external cycle at one week, and the arithmetic is 7 x 24 x 60 divided by 10, or roughly a thousand inner iterations before the slowest signal returns [1]. The exact figure follows from the assumption, but the ratio is the managerial content: whatever the outer loop would have caught arrives about a thousand turns after the agent started acting on it.
A skeptic will point at the byline. Jiao is chief executive of a company selling AI-powered testing tools [1], and the claim that models and orchestration are becoming cheap and interchangeable while the verifier remains the scarce, load-bearing piece is exactly the sentence that business needs to be true [8]. The mechanism can still be inspected on its merits. What the provenance should push to the back of the queue is the pricing claim, that inside a strong loop a cheaper model can close the gap with a frontier model on whether the software works [13]. An incentive is not a refutation, but it is a reason to read that particular paragraph last.
One detail deserves to survive the argument. Agents are becoming primary users of developer tooling, and an agent cannot click through a dashboard or glance at a chart while it can reliably run a command and read the result [14]. That is a tooling constraint with a budget attached, because the systems a loop can verify against are the ones that speak in text.
The decision available this quarter is narrow. Who owns the definition of done for each agent loop, and whether that definition lives somewhere it can be audited rather than in the habits of whoever set the loop running. The consequence arrives later, once those definitions have hardened into the infrastructure every team's agents are graded against, and amending one means renegotiating what shipping means.
