Agents learned to act. They still can't listen.

Watch an AI agent work in 2026 and you see a strange asymmetry. The acting side is solved: agents browse, file, book, transcribe, and call APIs in chains that run for hours. Then you read the trace and find where they spend their lives. They wait. They wait for a human to read a notification, open a keyboard, and type the next instruction. The most autonomous software ever built idles on the world's slowest input device.

That input device has not moved in decades. Typing runs near forty words a minute; speaking near a hundred and fifty. Both are serial streams through one narrow channel, and the channel has a measured size. Zheng and Meister, at Caltech, put the numbers in Neuron: the human senses take in on the order of a billion bits per second, while human behavior puts back out about ten. One millionth of what comes in goes back out. Every interface you have ever used sits on the wrong side of that ratio.

Agents got a foundation model for acting. The input side is still a keyboard.

Why now is different

Reading intention upstream of the fingers used to be science fiction with a lab budget. It isn't anymore. Each modality got its foundation model in turn: text, then images, then audio. Brain signals are next in line, and the first pieces are already public. MANAS-1, an EEG foundation model trained on sixty thousand hours of recording, shipped from India in February 2026. Meta's Brain2Qwerty line decodes typed sentences from brain activity, and the newer code is open. Conduit, Sabi and Mesotron are building products on the same premise.

What is still missing is not cleverness. It is data: brain signal recorded together with speech, at scale, in the languages people actually think in. Clinical archives never collected it, because their clinical questions were seizures and sleep, not language. That gap is a company-sized gap, and it is the one we chose: Nurosis collects the paired data itself, raises one model on it, and lets agents run on top of what that model decodes.

The shape of the stack

Three phases, each feeding the next. Phase 1 records EEG while people speak and keeps both streams, word-aligned. Phase 2 grows one foundation model from those recordings alone. Phase 3 is the part agents care about: the model becomes a base layer other software builds on, the way text models did. An agent attached to it does not wait for you to type. It hears the intention, asks for confirmation when confidence is low, and acts.

We are open about where this stands: thesis stage, hardware and study design being settled now, first recordings ahead of us. The direction is not in doubt; the timing is the work. When people ask what the next interface for AI agents looks like, our answer is short. It looks like no interface at all.