What an agent does with a decoded thought

A decoder that turns thought into text is half a product. Text on a screen still needs hands. The other half is an agent: software that reads the decode and does things with it, sending the message, filing the note, starting the task. This is the layer where brain-computer interface work meets the agent stack everyone is building, and it has its own engineering discipline. Three rules run it.

Rule one: the decoder speaks in probabilities, and the agent must listen

A decoder never says 'the word was water'. It says 'water, seventy percent; what, twenty; wonder, four'. A naive agent acts on the top guess and sends the wrong message. A correct agent holds a threshold: above it, act; below it, ask. The asking can be a whisper, a glance at a confirmation, a voice prompt, anything cheaper than an undo. When we describe approving a message by voice while driving, this is the machinery underneath: the thought drafts, the probability decides whether the draft is trustworthy enough to show, the voice approves. An agent that asks when unsure is not weaker than one that guesses; it is the only kind anyone will keep switched on.

Ask when unsure is not a limitation. It is the whole interface contract.

Rule two: speed is part of correctness

A confirmation that arrives late is a distraction, and a distraction at sixty miles an hour is a safety issue. The useful budget is the patience of a conversation: an agent that responds within the natural rhythm of speaking feels like a participant, and one that lags feels like a form. That budget has to cover the decode, the confidence check, and the rendering of the question. It is a systems problem, not a benchmark number, and it is why the decode and the agent are designed together rather than bolted together.

Rule three: the signal stays home

Raw neural data is the most personal data a person produces, and European law already files brain recordings under special-category protections that require explicit consent and tight handling. The architecture that earns trust is boring and local: decode on the device, transmit intent rather than signal, keep raw recordings in the room where they were recorded, and let the cloud see the errand, never the brain that ran it. Privacy by architecture costs engineering effort and buys the only thing harder to obtain: permission to exist at scale.

Put the three together and the agent layer stops being science fiction and becomes a quiet protocol: think, decode, weigh, act or ask, confirm, learn from the correction. Every correction is labeled training data. The interface gets better the more you use it, which is the rare kind of flywheel that respects the person turning it.