AI Agents: The Poltergeist in the Machine

The essay starts from the poltergeist, the "noisy spirit" that moves things, to describe how language models, which once only spoke, learned to act: opening files, running programs and chaining steps to carry out tasks. It traces the word agent from Latin agere and the legal figure of someone who acts on another's behalf, the old dream of artificial servants from Hephaestus's workshop in the Iliad to Čapek's robots, and the rational agents of Russell and Norvig with their loop of perceiving, deciding and acting. It explains how the model became an agent from outside, through a harness that executes its requests (ReAct, Toolformer, the Model Context Protocol), distinguishes a programmed workflow from a genuine agent, and follows AutoGPT, the simulated village of Smallville, computer use, and the difficulty of measuring progress with SWE-bench, METR's time horizons and MirrorCode. It closes with the principal–agent problem: the egoless phantom will not betray us out of ambition, but it can follow instructions to the letter without understanding what we wanted, so delegating requires defining success, verifying it, approving irreversible decisions and setting budgets.