The two primitives: Ground and Act.

Cognition only does two things: it knows, and it does. Every capability, every connection, every flow of information resolves into one of two primitives, Ground or Act.

Ground is everything that constitutes what the system knows and who it is. The most critical grounding signal is the human, not as a user issuing commands from outside, but as a participant inside the grounding layer, providing authority, direction, correction, and situational context. Alongside the human, Ground includes ambient awareness from daily work, a living working memory that maintains situational awareness continuously, a long-term memory that serves as the persistent identity store, adversarial synthesis from multiple frontier models, and the world itself through research and search.

Act is everything the system can do: the expression of grounded knowledge into the world. Communication, computation, perception, scheduling, building. Two parts of Act matter most. Every action is cryptographically attested, creating a tamper-evident record, so trust is structural rather than a policy bolted on afterward. And the system can propose changes to its own infrastructure, but it cannot approve them; a change passes through a human authorization checkpoint before it takes effect.

The insight that led here was practical. After building enough AI systems with twelve, twenty, thirty individual tools, the pattern became obvious: every tool was either a way of reading something or a way of changing something. Consolidating them into two primitives does not reduce capability. It clarifies the only decision the model actually needs to make, and pushes all provider complexity below the interface, where the model never sees it and where it can be swapped without breaking anything above.

There is a competing view that a capable enough model should construct its own tools at runtime. The research is real, and it holds when the generated tool becomes a stable artifact: created once, frozen, called repeatedly. It breaks when tool generation becomes a per-session activity, the model rebuilding its integrations every conversation rather than calling a known interface. That is where the consistency guarantees that make tools trustworthy disappear.

The same pattern reaches even the surface responsible for building and maintaining the system. A dedicated builder surface exists to externalize hands-on infrastructure work, so the cognitive entity does not absorb that load directly. It is not a separate architecture. It grounds itself in the shared memory substrate and the operator's direction, then acts on the infrastructure with human authorization gating every change. That even the builder exhibits the same Ground and Act pattern is the strongest evidence that the two primitives are not specific to one entity. They are universal.

The dominant paradigm treats AI as a stateless utility: it answers a question or executes a task, then forgets everything and starts over. Each interaction begins from zero, and the human carries the entire burden of context, what happened yesterday, what was decided last month, what the client said, why this path and not that one. The AI does the labor; the human does the remembering. That is not a partnership. It is a very fast tool operated by an increasingly exhausted person.

The two primitives invert this, because Ground, by design, is cumulative. The system carries its own context. It remembers what it was told, what it observed, what it did, and what happened as a result, and every grounding signal feeds back into the substrate, available to every future act. The human stops being a context-loader and becomes a director, providing judgment, correction, and authority while the system carries the growing weight of institutional knowledge.

The limiting factor in knowledge work was never the speed of execution. It was the bandwidth of human attention. Senior people are not slow; they are saturated. They carry more context than they can act on, and every delegation requires a context transfer, every new hire takes months, every departure takes knowledge with it. A system that grounds itself continuously and acts with attested accountability does not solve this incrementally. It changes the category. The context lives in the substrate: shared, searchable, continuous, and compounding.