Speakers at AI Engineer World’s Fair debate the limits of agentic automation and the role of human agency

Wednesday’s autoresearch track at the AI Engineer World’s Fair (AIEWF) highlighted a debate over how much autonomy agents should have in engineering workflows. Roland Gavrilescu of Introspection described autoresearch as an “outer loop” in which agents help study and maintain the primary “inner loop” of development, framing it as a collaborative process rather than a replacement for human work.

Anthropic’s Thariq Shihipar, speaking on Thursday, echoed the idea that models are “grown, not developed,” emphasizing continuous discovery and adaptation during use. This perspective aligns with autoresearch’s iterative, loop-based approach to system improvement.

Addy Osmani, former Google engineering leader, drew a sharper distinction: agents can run much of the inner execution loop, but the outer loop—setting goals, judging results, and taking responsibility—should remain human. He summarized this as “The inner loop is capability. The outer loop is agency.”

Notion’s Geoffrey Litt argued that metaphors like “software factory” risk depersonalizing engineering and sidelining human creativity. He warned that those who delegate understanding may be replaced by agents, while those who retain it will continue to drive innovation. Litt’s stance centers on the need for engineers to stay active participants in the creative process.

Paul Bakaus, creator of the design tool Impeccable, rejected both fully manual design and fully automated “loop-maxing.” He proposed a middle path: agents handle the laborious first 80% of work, while humans step in for the final 20% to apply taste and perspective. Bakaus insisted that “there is no auto, and there will be no auto,” emphasizing that authorship and responsibility require human involvement.

In generative media, Google’s Nicole Brichtova distinguished between average preference and cultivated expertise, arguing that human judgment is essential to elevate outputs beyond default aesthetics. She suggested closer collaboration between model developers and creative experts to ensure results reflect intentional design rather than inherited biases.

Adobe principal scientist Carlos Sanchez demonstrated “agentic sites” that assemble and personalize web pages in real time, but cautioned that without clear goals and constraints, agents can stray outside brand guidelines. He framed the challenge as knowing what to build, not just how to build it, reinforcing the need for human oversight in agent-driven systems.