Let me steelman the hype before puncturing it. "Agentic AI" is not pure marketing. There is a real, distinct capability behind the word: a model that plans, takes actions, observes results, and continues — rather than producing one answer and stopping. And it's protected IP. Anthropic's grant US12619815B2 (May 2026) describes a multimodal agent that automates image-text interfaces — software that reads a screen and operates it. That's an agent doing something, not a chatbot describing it.
Now let's check the story everyone's telling against what's disclosed. The popular claim is that agentic AI is here and reliable. What the patent actually covers is a method — a perception-and-action loop, with robustness tricks like the "magnitude invariant" perception in the grant's title. A method claim is not a reliability claim. The very fact that the patent emphasizes robustness to screen scale tells you brittleness is a known, hard problem the inventors were specifically working around.
I'd love to believe the agents-everywhere narrative, but the gap is structural. An agent that succeeds 70% of the time at a ten-step task fails most of the time end to end, because errors compound. The patents describe how to make each step work; they don't — can't — promise the chain holds. That compounding is the deflationary fact the hype skips.
So is "agentic AI" a new capability or a new word? Both, honestly. The capability is real and patentable, as the Anthropic family shows. The word is doing extra work, stretching a genuine-but-fragile mechanism into an implied dependability the disclosures don't support.
The grounded read: when someone sells you an agent, separate the demo from the duty cycle. The patent tells you the mechanism exists. It tells you nothing about whether it'll finish your ten-step task without supervision — and that, not the mechanism, is what "agentic" is being sold on.