AIEWF Daily Dispatch: The great loops debate and the state of AI engineering

The last day of the AI Engineer World's Fair closed with a debate that summed up the whole week: are AI coding agents — 'loops' — ready to run software development on their own, or is everyone just pretending they are? Moderator Allie Howe framed it bluntly, asking whether there's a gap between the hype around loops and what actually works in practice. An hour of arguing followed, and nobody in the room walked away with a clean answer.

On the optimistic side sat Geoffrey Huntley, who built the Ralph Loop, and Keycard CEO Ian Livingstone. Huntley's position was simple: loops are already here, and there's no going back to writing code by hand. Livingstone argued that verifiability is what actually matters, not the method that produced the code, and pointed out that iterative loops — try, learn, apply — have always been how software gets built. The only real change, he said, is speed. Dex Horthy from HumanLayer and Greg Pstrucha from Subroutine pushed back hard. Horthy wasn't against loops themselves; he noted Kubernetes runs on deterministic control loops too. His problem is that the hype has outrun the engineering discipline needed to make agentic loops safe to trust with real systems. Pstrucha added an economic angle, arguing that teams can't simply buy more tokens to make an unsustainable process work.

Anthropic gave the debate a concrete example to chew on. Mike Krieger, the company's Head of Labs and an Instagram co-founder in an earlier life, described Claude Tag, the internal tool Anthropic revealed last week. Engineers there aren't handing tasks to Tag one at a time — they're assigning it ownership of entire slices of the codebase and telling it to monitor feedback and act on its own. Krieger called this shift genuinely transformative for how his team works. But he also admitted the team is now bottlenecked on human review, and on people's ability to fully grasp what the agents are actually doing across the system.

Barr Yaron's annual Amplify survey put numbers behind the mood in the room. Ninety-five percent of AI engineers now use agents, roughly double last year's figure, and 89% of those said their agents can write data directly, up from 52%. Yaron's line was memorable: agents have moved past reading and summarizing and are now taking actions inside systems. Yet the safety net hasn't kept pace — human approvals remain the dominant control, with everything else a scattered mix of sandboxing and memory tricks. Forty percent of respondents said AI costs already limit their ambition, and 59% worry that the code being generated today is quietly building tomorrow's technical debt.

The closing keynotes swung back toward optimism, as closing keynotes tend to do. Theo Browne showed off projects that would have needed a startup a few years ago and now fit into a side project. Garry Tan of Y Combinator said the fastest-growing founders he sees treat AI as a workforce, not autocomplete, and told the crowd to build AI-native companies rather than companies that merely bolt AI on. It's a tidy pitch. Whether the discipline underneath it is ready is exactly the question the loops debate never quite resolved.