Quoting Andreas Påhlsson-Notini

The critique isn't just philosophical posturing: it points to a specific architectural problem. If agents are trained on human-generated data and evaluated by human raters, they will reproduce human cognitive habits by design, not by accident. That's a structural constraint, not a correctable bug.

This connects directly to the tensions surfaced in our coverage of OpenAI's Agents SDK update from April 15, where the emphasis was on developer tooling and sandboxed execution rather than any rethinking of how agents reason or decide. The Påhlsson-Notini critique implies that better harnesses don't address the underlying imitation problem. It also resonates with the MIT Technology Review framing from April 16 ('Treating enterprise AI as an operating layer'), which argued that competitive advantage lives in infrastructure governance rather than model behavior. If agents are fundamentally mimicking human indecision and constraint-negotiation, then the operational layer absorbs those flaws too, and the governance problem becomes harder, not easier. Our coverage of human oversight illusions in AI warfare contexts adds a sharper edge: systems that inherit human hesitation patterns while operating at machine speed may be the worst of both worlds.

Watch whether agent benchmark designers begin publishing evaluations that specifically test for non-human problem decomposition strategies in the next two quarters. If no major lab introduces such a benchmark, that absence itself confirms the field isn't yet taking this critique seriously.