A debate about "loops" was a highlight of the final day at the AI Engineer World’s Fair, and it encapsulated a central conference tension: are autonomous agent-driven "software factories" workable today, or is engineering practice falling behind the ambitions?
The session was moderated by Allie Howe of Keycard, who opened by asking whether there is a real gap between the hype around loops and what actually works in practice.
Arguments in favor of loops
Pro-loop speakers included Geoffrey Huntley, creator of the Ralph Loop, and Ian Livingstone, CEO of Keycard. Huntley argued that loops are already present and irreversible: "It’s inevitable, it’s here to stay," and added that he could not imagine returning to hand-writing code.
Livingstone focused on verifiability: the important question is whether code can be verified, regardless of how it was produced. He also noted that loops have always been central to software development: "A loop is at the core of ‘I try something, I learn something, I apply something.’ And all we’re really talking about is how quickly we can expedite that process."
Huntley offered an analogy for optimizing loops: "We’re kind of like locomotive engineers now. That’s our job: to keep the locomotive on the rails."
Skeptical perspectives
On the skeptical side were Dex Horthy of HumanLayer and Greg Pstrucha of Subroutine. Horthy emphasized he is not anti-loop — he pointed out that Kubernetes is built on control loops — but those are deterministic. His concern is that "the hype is outrunning the discipline."
Horthy said he has not seen proof that the field is ready to step up an abstraction level to agents that fully control coding; if anything, he suggested the need to step down a level of abstraction.
Pstrucha raised economic concerns: he questioned the sustainability of agentic loops and warned against the idea that you can "orchestrate your problems away by buying more tokens."
Software factories and practical risks
The conversation moved to the "software factory" metaphor that has gained traction in the industry. Horthy warned that in a fully automated, factory-like agent environment engineers "never touch the problem," and recommended starting small and iterating with agent loops to build intuition rather than attempting end-to-end automation from the outset.
Even Huntley acknowledged dangers: he said software factories represent a likely future direction but cautioned the market has not yet solved the approach, calling it "frontier thinking."
The session ended with Howe polling the audience on which side had won. Ironically, the vote was frustrated by human error: stage lights were too bright for Howe or the panel to see how many hands were raised.
Anthropic and Claude Tag: an example of delegated workflows
Anthropic may illustrate a move toward a software-factory model. Mike Krieger, Instagram co-founder and Head of Labs at Anthropic, spoke in a morning session with swyx about Anthropic’s internal model, Claude Tag, which the company announced the previous week.
Krieger described Claude Tag as more delegated, asynchronous and proactive than Claude. That suggests a practical early software-factory pattern: not agents outright replacing teams, but multiple people delegating responsibilities to a system like Claude Tag.
He said most usage is "much more delegated," and gave an example of instructing agents: "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." He said that approach "has really changed how we operate currently," moving to a multiplayer, async, proactive model.
Krieger also noted downsides: his team is "bottlenecked on reviews" and limited by "human ability to fully conceptualize what we’re doing."
2026 AI Engineer Survey: agent adoption surges; controls and costs lag
Barr Yaron of Amplify presented the company’s annual industry survey. According to Amplify, 95% of respondents now use agents — roughly double last year’s share. Among teams using agents, 89% said those agents can write data in systems, up from 52% the previous year.
"Agents are no longer reading, summarizing, drafting," Yaron said. "They’re taking actions inside the systems."
Despite broader adoption, control mechanisms remain comparatively primitive. Human approvals and permissions were the two leading safeguards, followed by a scattered set of task decomposition, retrieval, memory and sandboxing techniques. "Nobody has settled the control layer for agents," Yaron said.
Cost is also constraining: 40% of respondents said AI costs regularly limit how ambitiously they use AI, and another 36% said it sometimes does. Token usage is now the second-most monitored production metric after quality.
The survey captured the conference’s central contradiction: AI has made experimentation cheaper and helped teams produce more software, but 59% of respondents fear that today’s AI-generated code creates long-term liabilities.
Closing keynotes: optimism about building with AI
The conference’s final talks returned to optimism about building with AI. Theo Browne showcased several projects he has built or is building with AI, arguing the scale an individual developer can realistically attempt has shifted: "What used to be a startup is now a side project," he said, and projects once considered "too big" are becoming reachable.
Garry Tan, president and CEO of Y Combinator, framed that optimism organizationally: the fastest-growing founders YC observes "are not treating AI as autocomplete, they’re treating it as a workforce." His closing advice was: "Build an AI-native company, not a company that just uses AI."
The week’s debates highlighted how much engineering work remains before an AI-native vision is broadly viable. The closing keynotes, however, reminded attendees why many engineers are pushing forward: they want to take control of these powerful systems and ride the locomotive into the future.
