On the closing day of the AI Engineer World's Fair, industry leaders debated whether loops are ready for mainstream use and whether humans still have a place in coding.
On the loops-now side, Geoff Huntley, founder of Latent Patterns, was understandably bullish. He said that over two years ago he was at Canva and watched engineers prompting repeatedly and realized that this could be done much more efficiently with a software loop.
"It's somewhat inevitable," he said. "It is not a complete silver bullet. This time next year, at the conference we're going to see a whole bunch of talks saying our factories fail, our loops fail. These are things that we are yet to figure out."
He likened the situation to the first deployments of Kubernetes, where it took the industry years to get deployment right. But when they did, it was a revolution.
Dex Horthy, CEO of HumanLayer, was less enthusiastic. Earlier in the week, he gave a keynote in which he described how — as an experiment — he'd taken humans out of the coding loop and let machines do the job. He monitored the results, and they weren't good. He felt it was clear that AI wasn't up to the job yet.
"I think the basic take here is not whether loops are good or bad," he said. "I've seen lots of people try to apply AI to this problem of having to review bots, but it doesn't feel to me like it's working."
The increasing use of loops was inevitable, he said, but too many people are rushing into the space with too little thought. A more measured approach was needed, and some of the most enthusiastic early adopters would get their fingers burned.
The debate, done in the Oxford-style timed format, was closely fought, and the audience was asked to vote on either side. Unfortunately, the stage lights meant the moderator couldn't count votes, but — based on our view of the audience — Horthy's viewpoint won by a narrow margin.
Top comments (4)
The Kubernetes comparison cuts both ways, and I think that's the honest part. Yes, k8s eventually won, but the years in the middle were full of teams running it because it was the new thing, not because they had the scale to need it, and paying for that in outages. Horthy's "too many people rushing in with too little thought" is describing exactly that phase. So both can be right at once: loops are probably coming, and most early adopters still get burned. The interesting question the panel skipped is what the safe on-ramp looks like, since with k8s it was managed control planes hiding the hard parts.
Counterintuitive take: loops aren't the bottleneck; architecture is. The future isn't 'fewer loops,' it's smarter orchestration and safer abstractions that bake looping efficiency in, while devs write higher-level intent.
the vote splitting basically 50/50 feels right because both sides are describing the same missing piece from different ends.
i run loops on real client work every week and the thing that decides success isn't the model, it's whether the loop has a referee it can't argue with. a test suite. a regression diff that must come back byte-identical. a scanner that fails the run if a banned word shows up in the generated output. this week i had an automated pass that validates a parser on one $0.01 probe before it's allowed to spend $20 on the full batch. the probe failed, the loop stopped, the budget survived. the expensive step was gated by a cheap external check, not by the model's opinion of its own work.
the loops that burn people are the ones grading their own homework. the model writes the code, the model reviews the code, the model declares victory. you removed the only adversarial element in the system and then blamed the model for agreeing with itself.
so huntley is right that it's inevitable and horthy is right that most people aren't ready, because most people are shipping the second kind of loop and calling it the first. different domain, same spine: nobody debates whether CI is ready, because CI never asks the code how it feels.
AI coding loops are promising, but they're not a replacement for engineering judgment yet. The real opportunity is human-in-the-loop systems where AI handles iteration and humans provide context, validation, and critical decisions. Adoption seems inevitable—but disciplined implementation will matter more than speed.
Even though it is just fulfilling the task with quick iterations.