On Wednesday at the AI Engineer World’s Fair main stage, autoresearch — loops in which agents observe and maintain systems — was a central topic. Speakers explored how much of development and creative work should be delegated to autonomous agents and how much must remain under human engineering control.
Roland Gavrilescu, co-founder of Introspection, described autoresearch in a morning interview as enabling loops where agents help maintain the system itself. He characterized autoresearch as an “outer loop” that studies and sustains the primary inner loop.
Anthropic’s Thariq Shihipar, who works on Claude Code, did not use the term autoresearch explicitly but echoed its spirit in his keynote: models are “grown, not developed,” he said, arguing that teams learn and discover through use as models evolve.
Who should hold the outer loop?
A recurring tension at the event was whether agents should control the outer loop or humans must retain that role. Former Google engineering lead Addy Osmani framed the distinction sharply: agents can handle much of the inner execution loop, but the outer loop should remain engineering performed by humans. His shorthand: the inner loop is capability; the outer loop is agency.
This debate pushed back against the conference rhetoric of fully automated “software factories” that had been prominent earlier.
Design and authorship: a middle path
Paul Bakaus, presenting his new design tool Impeccable, rejected both extremes of entirely manual design and fully hands-off automation. He described a practical middle path: let agents handle the laborious first 80% of the work, then bring humans in for the final 20% to add taste and a unique point of view.
Bakaus also tied this stance to authorship and responsibility. He argued people need purpose and want a role in what they create, and that working with an agent increases ownership of the product. Impeccable, he said, will never one-shot a finished solution — users must participate and be able to steer the outcome. As he put it: “There is no auto, and there will be no auto.”
Generative media: whose judgment shapes outputs?
Panels on generative image, video and audio models raised a related question: beyond raw capability, whose sensibility determines the result? Nicole Brichtova, who works on Google’s generative media products including Nano Banana, drew a distinction between average preference and cultivated expertise: practiced creators notice things that average users do not.
Brichtova warned that every model carries a default aesthetic, often reflecting the modeling teams’ choices: “It ends up being us.” She suggested developers should collaborate with people who have a strong creative point of view, effectively bringing an art director back into the loop.
Shane Gu made a similar point: even as models can generate and refine content, humans must keep the sensitivity to detect when results are wrong, generic, or insufficient.
Agentic sites and brand control
The web itself is confronting automation. In a session on “agentic sites,” Adobe principal scientist Carlos Sanchez showed websites that assemble and personalize pages in real time based on a visitor’s intent. Sanchez presented this shift as inevitable: the capability exists and will get better, cheaper and faster.
He also cautioned that while AI makes it easy to build, it can be hard to know what to build. That matters particularly when an agent generates experiences on behalf of a brand: a fully generated site can stray from brand guidelines, so you cannot simply hand over the entire site to automation without safeguards.
Conclusion: automation grows, human responsibility remains
The day’s discussions made clear that autoresearch and agentic technologies are maturing and compelling, but most speakers argued humans must stay in the loop at the level of goals, judgment and responsibility. Agents may observe, evaluate and even improve other agents, yet people still need to define objectives, assess outputs and accept responsibility — and preserve authorship and taste.
The likely outcome is hybrid workflows: agents perform repetitive or laborious tasks at scale, while humans retain the substantive decision-making, creative judgment and accountability.
