Why AI loops stall: the hidden cost of chasing a target that can't be reached
Knowing When to Stop: The Art of Making a Loop Converge
Loop engineering promises to let AI agents work autonomously, but a loop is only as good as its verifier. Drawing on interviews and experiments, this piece breaks down the four conditions for convergence—target state, observable state, precise edits, and a stopping rule—and shows why code loops thrive while open-ended visual tasks stall. The economics are brutal: returns are logarithmic, and past a plateau, extra iterations can actively hurt. A loop is tuned to its stack, and bespoke loops don't generalize for free.
Loop engineering is not the practice of making an agent retry. It is the practice of making each cycle reduce the distance between the current state and a desired state. A loop is not yet a direction.
- ausbah
i’m not trying to dismiss this article entirely as i did get value around from discussion around verifying
but this coming from a VC firm invested in token producing companies and quoting the “genius” behind openclaw makes me think this is just an psyop of over engineering to get ppl to blindly spend tokens on the LLM slot machine
- Sha1rholder
I failed to think of any polite words to express my feelings about this article.
- intended
There were A16z articles I remember reading and being better off for. This… I struggle to… separate the managerial crack from what insights remains of a field of engineering.
It isn’t helped that at random sections it feels like I’m reading Claude speak. Every instance of which the article is better off by its excision.
The combination of statements that sound like voodoo, followed by the articulation of something obvious (“you need to have a way to encode what good is.”), is disheartening to say the least.
At one point I wondered if this article was actually being subversive; attempting to bring back rationality by stating the need for Loops to be constrained by economic limits
> “ The missing piece is boring yet necessary: cost per iteration, progress per dollar, a curve someone can see while the loop is still running.”
Where engineering used to be, a religion seems to be standing. Opinion and positions instead of measurement and clarity.