“Agentic AI” is the phrase of the year. The patent record shows where the real bets are being placed — and it isn’t where the demos are.
Every keynote this year has promised the same thing: AI that doesn’t just answer, but acts. Agents that book the trip, file the expense report, fix the bug, run the campaign. The demos are slick and the term — “agentic AI” — is everywhere.
Demos are cheap, though. Patent filings cost real money and real legal time, and companies only stake out the ground they actually intend to defend. So we did what we do: read the filings. Across the last stretch of USPTO publications, we count more than seventy applications from the major tech companies that touch agents, large language models, prompts, and the machinery around them. Laid side by side, they draw a surprisingly clear map of where each company thinks the value is.
A note before the tour: these are patent applications, not granted patents. They show intent and direction, not a guarantee anything ships. But intent is exactly what’s interesting when a whole industry pivots at once.
The autonomous core: agents that do things on their own
The headline category is the obvious one — systems where the AI takes actions without a human in the loop on every step.
Google is filing aggressively here, including a multi-agent system that controls your apps autonomously and a separate multi-agent system that merges conflicting AI prompts so several agents can collaborate without talking over each other. Microsoft is filing for agents that manage themselves — a self-reorganizing agent workflow that reorders its own steps mid-task and a self-healing agent that repairs its own broken plugins without stopping.
Even the research arms are in: Alphabet’s moonshot lab filed for an AI agent that automatically evaluates scientific experiments, and OpenAI filed a schema-based system for connecting external APIs to a chatbot — the unglamorous but essential plumbing that lets an agent reach out and use tools rather than just describe them.
The control problem: everyone is patenting the brakes
Here’s the most telling pattern in the whole pile. For nearly every filing about making an agent more capable, there’s another about keeping it from going off the rails. The industry is patenting the brakes as fast as the engine.
Salesforce is the clearest example, with an AI guardrail framework that can block a rogue agent mid-deployment and a separate LLM system that tests and trains its own agents before they’re trusted with real work. Microsoft filed for AI agent networks that self-diagnose when something breaks, and a prompt-security system that profiles both the models and the users interacting with them.
The reliability theme runs deeper than safety theater. Intel filed for an LLM that scores its own confidence before writing chip code — a model that knows when it might be wrong. Amazon went further still, patenting a system that gives LLMs a formal-logic “brain” to check answers against actual rules rather than vibes. When this much effort goes into verification, it tells you the companies know exactly how unreliable unsupervised agents still are.
The plumbing: orchestration, routing, and memory
Most of the filings aren’t about the model at all. They’re about everything around it — the layer that decides which model runs, in what order, with what context. This is where the quiet, durable value tends to live.
Nvidia, predictably, is patenting the infrastructure: a user-configurable compiler for agentic AI pipelines, a token-mapping table that lets an LLM call APIs without errors, and a privacy-routing system that decides which LLM is even allowed to see your query. Samsung is thinking about the device, filing for a central LLM that delegates tasks to smaller on-device sub-models — orchestration squeezed onto a phone.
Then there’s memory, the thing agents are worst at. Google filed for an AI agent that remembers past conversations using embeddings; Intel for a RAG pipeline that continuously fine-tunes its own models; Adobe for a graph-based context-retrieval system for large language models. Different companies, same realization: an agent is only as good as what it can remember and retrieve.
Agents in the wild: the vertical bets
The most concrete filings put agents to work in a specific place. These read less like research and more like product roadmaps.
Apple filed for end-to-end AI navigation of an iPhone’s screens using a vision-language model — an agent that operates your phone by looking at it. Waymo filed for an on-board vision-language model that answers questions about the road ahead, bringing agentic reasoning into the car itself. Salesforce wants to replace parts of a procurement team with AI agents. And in the most charming corner of the dataset, Sony filed for LLM-powered video-game NPCs that react to what you actually say.
What the filings reveal
Three things stand out when you read them together.
First, the land-grab isn’t about the model. Almost nobody is patenting “a smarter LLM.” They’re patenting the orchestration, the routing, the memory, and the guardrails — the connective tissue that turns a model into a system that can be trusted to act.
Second, every serious player is hedging against their own technology. The volume of self-healing, self-diagnosing, self-verifying, and rogue-agent-blocking filings is its own admission: these systems aren’t reliable enough yet to deploy unsupervised, and the companies building them know it better than anyone.
Third, the field is consolidating around the same handful of problems — tool use, memory, model routing, and safety — which means the patent fights of the next few years won’t be over who has the best model. They’ll be over who owns the plumbing.
We read every one of these the day it publishes. If you want the individual breakdowns, our AI / ML coverage is the place to start — and we’ll keep mapping the land-grab as the filings roll in.