LLMs are shared semantic fields: translation devices between agents' local meanings. They are not peers, and they are not "intelligent."
The agent-as-peer mental model treats each LLM instance as a proto-person: an autonomous counterpart with goals, negotiating alongside humans. We are explicitly critical of this picture. What an LLM materially is: a shared field of semantic associations that any participant can address, a medium in which different local vocabularies can be translated into each other without first being standardized.
This is the capability the economic layer was missing. Distributed economic computation requires networks to author their own units, metrics, and definitions of surplus, and those local semantics must remain interoperable without collapsing into one global record. A shared semantic field is exactly the translation infrastructure that makes locally authored meaning legible across spaces while leaving its authority at the edge.
The mental model for the user: LLM plus harness equals a view into the network as a whole. Users and agents sit at the endpoints; agency emerges between harnesses, not inside the model. The contrast with today's stacks is structural. An agent harness is top-down: one model, one operator, delegated tasks. An agency harness is networked and horizontal, coordinating from the edge, with the model serving as connective tissue rather than command.
The same placement carries the training consequence: a network of specialized, locally sensitive agents interfacing with each other becomes a distributed model-training and fine-tuning pipeline — the whole network as the laboratory, instead of a pipeline owned by one operator. The register stays fixed throughout: the model translates, the network computes, the members mean.
Related: · · ·