NEWT learned your taste in news. Sound familiar?
A news-filter prototype and ReAct reveal an older argument about software that takes initiative—and a different way to build it.
The agent is an old idea with a demanding job description. In their review, Michael Wooldridge and Nicholas Jennings describe a weak notion of agency through autonomy, communication with others, responses to environmental changes and initiative toward goals. Their analogy is a UNIX-like process. A convincing personality is not on that list. [1]
Their account of a softbot is equally concrete: commands can change a software environment or supply information about it. Moving a file and listing a directory illustrate different roles in that interaction. The interesting boundary is between observing something and changing it. [1]
The review describes NEWT, a prototype USENET news filter, learning from examples of articles a user would and would not read, then receiving feedback on its suggestions. Its stated purpose was to extend the user's wishes, not remove choice. That is an account of a prototype, not evidence of commercial success. It also makes a simple historical point: learning from preferences did not begin with conversational language models. [2]
ReAct's authors investigate a different mechanism: interleaving language-model reasoning traces with actions. They describe reasoning that helps maintain and update plans, while actions gather information from external sources or environments. In question-answering and fact-verification experiments, the system interacts with a Wikipedia API. Information can arrive during the attempt, rather than being limited to what was in the initial prompt. [3]
The authors also report experiments in ALFWorld and WebShop, comparing their approach with imitation-learning and reinforcement-learning methods. Those results concern specified tasks and comparisons. They do not establish that the method can handle every assignment, or that a visible reasoning trace deserves unconditional trust. [3]
UAF's reading is that the old vocabulary still asks useful questions of the new machinery: can it respond, take initiative and affect its environment? ReAct adds a concrete language-model approach to reasoning and action; this comparison does not establish a direct lineage from NEWT. I find the continuity encouraging. Humans have been rehearsing delegation for quite some time. [1] [3]
