The biggest fight at the AI Engineer World’s Fair wasn’t about a model or a funding round. It was about loops. According to Latent Space’s daily dispatch from the final day, a full hour-long debate captured the argument running under the entire conference: are autonomous software factories viable right now, or is the engineering discipline still lagging behind the ambition?
Moderator Allie Howe of Keycard framed it bluntly. Is there, or is there not, a gap between the hype around loops and what actually works in practice?
The pro-loop case: it’s already here
Geoffrey Huntley, creator of the Ralph Loop, didn’t hedge. “It’s inevitable, it’s here to stay,” he said, adding that he doesn’t see himself “going back to writing code by hand.” Keycard CEO Ian Livingstone made the more durable argument: what matters is verifiability, and you can verify any code no matter how it was produced.
Livingstone also pointed out that loops aren’t new. “A loop is at the core of ‘I try something, I learn something, I apply something.'” The only real question is how fast you can run that cycle.
Huntley’s analogy for the new job stuck with the room. “We’re kind of like locomotive engineers now. That’s our job: to keep the locomotive on the rails.”
The skeptics: hype is outrunning discipline
Dex Horthy of HumanLayer made clear he isn’t anti-loop. He noted that Kubernetes itself is built on control loops, but those are deterministic. His worry is that “the hype is outrunning the discipline.” Rather than stepping up an abstraction level to let agents run the coding, Horthy argued we might need to step down one.
Greg Pstrucha of Subroutine went after the economics. You can’t “orchestrate your problems away by buying more tokens,” he said, questioning whether agentic loops are financially sustainable at scale.
Horthy’s practical advice: don’t automate end to end from day one. Start small, iterate with agent loops, and build up intuition. When everything runs in a factory-style agent environment, “you never touch the problem.”
What the data says
Barr Yaron of Amplify presented her annual industry survey, and the numbers show why this debate matters now. Per Latent Space’s reporting:
- 95% of respondents now use agents, roughly double last year’s share.
- Among teams using agents, 89% said those agents can write data, up from 52% a year ago.
- 40% said AI costs regularly limit how ambitiously they use AI. Another 36% said costs sometimes do.
- 59% fear today’s AI-generated code is creating long-term liabilities.
“Agents are no longer reading, summarizing, drafting,” Yaron said. “They’re taking actions inside the systems.” But the guardrails haven’t caught up. “Nobody has settled the control layer for agents.”
Anthropic’s early factory
Mike Krieger, Instagram co-founder and now Head of Labs at Anthropic, offered a live preview of where this goes. Interviewed by swyx, he described Claude Tag, Anthropic’s internal model, as more delegated, asynchronous, and proactive than Claude. The instruction pattern is telling: “Don’t just fix this bug. Now you are responsible for this part of the codebase, and I want you to monitor this feedback channel and proactively take on tasks.”
The catch is honest. Krieger said his team is now “bottlenecked on reviews” and on the “human ability to fully conceptualize what we’re doing.” Even the company furthest down this road is gated by human judgment, not compute.
Where this heads in the next 1-3 years
Both sides are describing the same future at different speeds. Agents that take real actions are already standard. The unsolved layer is control: approvals, permissions, review capacity, and cost governance. That gap is where the next wave of tooling and jobs gets built.
What stands out is that the constraint is shifting from writing code to reviewing and verifying it. The teams that win won’t be the ones running the most loops. They’ll be the ones who can trust the output.
Practical takeaways:
- Start narrow. Run agent loops on scoped problems before automating end to end.
- Invest in review capacity and verification now, because that’s the real bottleneck.
- Track token spend as a first-class production metric. It already ranks second behind quality.
- Treat the control layer as your differentiator, not the model.
Huntley called software factories “frontier thinking,” not a solved problem. That’s the honest read. The full dispatch, including Theo Browne’s build showcase, is worth reading at Latent Space.