Electricity is not a product. It is drawn on everywhere, and it is dangerous to touch. Nobody who works with it does so bare-handed.
The same shift is under way with what most people call AI. It is becoming a general-purpose substrate: a new kind of electricity, drawn on everywhere rather than used in one place. And most of what is being sold on top of it is appliances — things you plug in and hand a job to. A chatbot. An assistant. An agent that goes off and does something in your name.
We think the useful question is not which appliance to buy. It is how a person works with the current itself, directly and safely, and directs everything plugged into it.
Our answer has a name: a cognitive companion. It is a new category, and it is not an appliance. It is the part of you that is wired in.
The definition
We use the definition the DAX Foundation publishes:
A cognitive extension of a person — part of them — through which they work with nonbiological intelligence and direct it. It is not a separate agent acting in their place.
Nonbiological intelligence is the Foundation’s term for what is usually called artificial intelligence. It names the substrate and nothing else, where artificial smuggles in a verdict that the thing is counterfeit.
Three things follow from the definition.
It is part of you. A companion is not a second party. You and your companion are one unit: its context, its trail and its direction are yours, and it has no interests of its own to negotiate with. In the Foundation’s words, it answers to its person, within a constraint declared in advance, and to no one else.
It is how you direct agents. Agents are going to do more and more of the work. Somebody has to tell them what to do, hold them to it, and check what they did. A companion is the part of you that does that — the hands, eyes and ears you work through when the work is being done by systems that are not you.
It is not an assistant, and it is not an agent. An assistant is a second party you hand tasks to. An agent acts in your place. A companion is neither. It is part of the one handing out the tasks.
More intelligence, in a precise sense
The Foundation defines intelligence narrowly, as the navigation of constrained possibility space: the how of getting from A to B using what is known. On that definition a companion gives its user more intelligence. Not more facts, and not a second mind, but more reach across what is possible, inside the constraints that apply.
That is the whole claim, and it is deliberately modest.
Where ours came from
This is not a category we arrived at from a whiteboard. We arrived at it from having built the thing and then finding that none of the available words described it.
In August 2021 we started building a personal AI agent, and published a video the same day setting out both what it would make possible and what it would cost us. GPT-3 was in limited beta. ChatGPT did not exist and would not for more than a year. That agent was called DAX.
For four years it got built, used, rebuilt and used again — not as a product for other people, but as the equipment one person needed in order to do the next piece of work. The compounding is the part that matters: each tool made the following one cheaper to build, so the early ones look like nothing much and the later ones look like cheating.
Along the way the word agent stopped fitting. An agent acts in your place, and this did not. It was not a second party at all. It was the part of the person that was wired in — holding the context, keeping the trail, and directing the agents that did the work. So we stopped calling it an agent.
It is worth saying plainly that we did not build the whole stack. The core is open source — OpenClaw, or Nous Research’s Hermes Agent, pick your poison — and we built everything it does not do around it. Have we built a personal agent? Of course we have. Everyone building seriously in this space has. The interesting part was never the agent.
What we are building
A companion in the general case is a person’s whole working life, which is a great deal of surface area and not a product.
What Obsidian Delta is working on is the engine: the cognitive companion itself, and the cognition it runs on. That makes us the cognition provider, and the division is deliberate. It is the same one we apply everywhere: the companion belongs to its person, who directs it and answers for what it is directed to do; the cognition underneath is our job, and how it is built and run is our liability. In the Foundation’s terms we are its operator, never its principal.
Products will be built on it. Each will be described in its own terms.
What happens when two of them meet
Here is the problem we did not have five years ago.
One person and one companion is a closed loop. The context is theirs, the trail is theirs, and the only party who has to trust any of it is them.
Now put two of those units in a room — two people, two companions, two organisations — and have them do business. Each side is acting through equipment the other side cannot see, on context the other side does not hold, at a speed neither person can audit in the moment. Nothing about that is sinister. It is simply a shape in which, afterwards, there is no durable record of what either side was actually bound by at the time.
That is the other half of this work, and it is why we did not stop at the companion.
Why we built an accountability layer to go with it — and what we intend it to compound — is in The Constraint. The protocol itself is at daxfoundation.org/#cp, published in full and openly licensed, as a working draft with one operational run behind it and no pretence otherwise.
More on this. It is the thing we are most interested in and the least finished.
Name the thing being done
There is a naming habit in this too. Nobody watching someone play music on a phone calls them an electricity user; they say the person is listening to music. Calling everything “AI” is the same category error. It names the substrate instead of the activity, and it drags the conversation into good-versus-evil before anyone has said what is actually being done.
So on this blog we will try to name the thing being done. When we write about a companion, we mean part of a particular person: the part through which they work with this substrate and direct what runs on it. Not a chatbot, not an assistant, not a product category, and not a stand-in for the person.
What comes next
This is the first piece in a series on cognition and the tools around it. Later posts will take up the cognition provider in its own right, what happens when companions start talking to each other at scale, and where our own work actually stands as it moves.
Where something is still an idea, we will say so. Where it has run once, we will say that too, and we will say how many times.