The job isn’t disappearing. It’s changing shape. Andrej Karpathy’s AutoResearch project (https://lnkd.in/gYt8RvX7) is a useful lens - though framing humans as “meat computers” misses the point.

The core idea is straightforward: instead of a researcher manually running experiments, an AI agent autonomously executes bounded trials. The human sets objectives and evaluates outputs. Karpathy calls this an “agentic loop”. Azeem Azhar describes the broader shift as the “loop economy”. Different labels, same pattern: the role moves from executor to evaluator.

That distinction matters. Execution is about throughput. Evaluation is about judgement. The former is easier to automate. The latter compounds with experience and is far harder to replicate. The people best positioned for this shift aren’t those chasing every new tool, but those with developed taste - the ability to rapidly distinguish strong outputs from weak ones, and explain why. That capability has rarely been valued in isolation. It will be.

There is, however, a clear downside: fewer evaluators are needed than executors.

Recent ‘Big 4’ narratives reinforce this trajectory - removing “scaling friction” attributed to humans by stacking loops. Agents supervising larger fleets of agents, with a single human overseeing the system ("human on the loop").

When humans start being framed as friction, the signal is unambiguous.