What AI Agents Actually Mean: Product Claims vs. Skepticism

In mid-May 2026, the discourse around AI agents splits sharply between impressive product claims and terminological skepticism. • Google unveiled "Magic Pointer," a Gemini-powered cursor that interprets intent from vague gestures, framed as a step toward ambient intelligence [1]. • Genspark claims $250M ARR in 12 months as concrete evidence that agentic AI delivers real productivity value [2]. • Simultaneously, Simon Willison amplified Boris Mann's argument that quantifying AI agents — "I use 11 agents" — is as meaningless as saying "I have 11 spreadsheets" [3]. The thread captures a live fault line between builders showcasing growth and critics questioning whether the language itself carries any signal.

If "AI agent" remains an undefined marketing category, it becomes nearly impossible for investors, users, or regulators to evaluate which claims reflect genuine capability versus polished framing. The gap between ambient-intelligence demos and a rigorous definition of autonomous work is where most of the real risk — and most of the real value — will eventually be discovered.

Two distinct camps are driving the AI agent conversation in May 2026, and they are largely talking past each other. On one side, companies are advancing specific product claims and revenue milestones that treat "agents" as a solved, deployable category. On the other, a quieter critical thread is questioning whether the word itself communicates anything at all.

Google's contribution to the first camp is "Magic Pointer," a feature that uses Gemini to interpret what a user is pointing at on screen — enabling references like "this" or "that" without a typed prompt [1]. Gemini Intelligence for Android extends this to app automation, web summarization, form-filling, and custom widget creation [1]. The coverage from The Neuron's Grant Harvey frames these announcements as a landmark shift: the interface begins to carry part of the prompt for the user, moving toward ambient intelligence where screens and keyboards are no longer the primary interface to computing [1]. Separately, Thinking Machines Lab previewed interaction models that process audio, video, and text in 200-millisecond chunks, enabling real-time listening and tool use [1], and Perceptron released a model that understands video as a stream of events rather than discrete screenshots [1].

Genspark offers a revenue-grounded version of the same argument. The company reports growing from $0 to $250M ARR in 12 months, with the final $150M added in roughly three months [2]. Its CEO's framing is conceptually tidy: LLMs are "brains without arms and legs," and agents are what you get when you give those brains tools, memory, and access to the software where work actually happens [2]. A live demo of its "Workspace 4.0" — researching a VC's preferences and generating a customized pitch deck in seconds — is offered as a concrete illustration [2]. Genspark also claims 100% of its own code is now written by AI, enabling small teams to ship at the speed of a single developer [2].

Against these narratives, Simon Willison amplified a terse critique from developer Boris Mann: saying you use "11 AI agents" is no more informative than saying you have "11 spreadsheets" or "11 browser tabs" [3]. The point is not that agents are useless, but that counting them strips out every piece of context that would make the claim meaningful — what the agents do, how reliably, at what cost, and to what end [3]. This definitional skepticism sits in direct tension with the growth metrics and interaction paradigm claims from Google and Genspark, neither of which addresses the underlying measurement problem.