The agent
The selection may be deterministic or stochastic. The act, once selected, executes deterministically because the substrate is a Z₁ machine. The agent does not produce the effect — the substrate does. The agent’s contribution is the choice, not the execution.
The role admits any selector
A human at a terminal is an agent. A cron daemon firing a scheduled job is an agent. A make build resolving a target is an agent. An init system starting a service is an agent. A language-model runtime emitting a tool call is an agent. The role admits any selector.
Each of these selects from the same substrate’s permitted-act set. Each pays per act in the same joules. Each leaves a record on the same clock. What varies between them is fluency, throughput, the selection process, and the cost basis — but not the agency.
Two species
Two kinds of agents currently address general-purpose computers at scale. Both occupy the same seat. Both exercise the same conferred agency. The substrate does not distinguish between them.
Wetware-substrate selectors. Trained by education, apprenticeship, and lived experience. Selection process is genuinely stochastic — wetware noise, attention, mood, context — and not reproducible even by the same agent under the same conditions. Throughput is low: one act per second of typing, one decision per second of attention. Cost basis is time. Fluency varies enormously across individuals.
Silicon-substrate selectors. Trained by programming (deterministic emitters) or by gradient-based optimization over a corpus (stochastic emitters). Selection process is either fully deterministic — compilers, schedulers, build systems — or sampled-stochastic, drawn from a learned distribution via a controlled-noise channel. Throughput is high: 10² to 10⁶ acts per second depending on the daemon. Cost basis is per-act energy.
The within-species variable
Fluency is not a property of the substrate. It is a property of the occupant. It can be measured as the fraction of the permitted-act set the agent can address, weighted by how readily it does so. Humans become more fluent through education; software daemons become more fluent through training or extension. The substrate’s permitted-act set does not change. Only the occupant’s reach into it does.
100%
substrate · permitted-act set
The ceiling. The total enumerable set of acts the substrate exposes.
~70%
Unix-fluent sysadmin
Decades of practice and a one-line pipeline that a non-technical user could not produce in a day.
~50%
Language-model runtime with shell access
A large reach across the act-set, mediated by sampled selection and the tool surfaces available to the runtime.
~25%
cron + make + systemd
Deterministic daemons addressing a substantial but bounded fraction of acts via scheduled timers, build rules, and unit files.
~5%
Non-technical user
A small reach into the permitted-act set through application interfaces — the act-set is the same; the fluency is not.
The history of computing is, in large part, the history of fluency expanding across occupants. The substrate has not grown more capable. The fluency profile of its occupants has rotated.
Continue