⟨ H, D ⟩ durable artifact
Authoring a Z₁ artifact
A human writing code, a shell pipeline, a SQL query, a regex. The authoring act is paid once. Subsequent executions run at the substrate’s thermodynamic floor.
The Periodic Stack of Agency
The seat is older than the occupant.
A computer permits a finite set of acts. The agent is whoever selects from that set. Agency is conferred by the substrate, not the actor. The seat has had many occupants — human computers, operators, programmers, schedulers, daemons, services, and now language-model runtimes — and only one ceiling.
act = ⟨ A, σ, p, t ⟩
The act coordinate · actor, selection process, primitive, timestamp
The substrate confers the agency
Conferred above, paid below.
A computer is a math machine. Voltage transitions across transistors, organized into Boolean gates, composed into arithmetic units, memory hierarchy, and I/O. The substrate does math. Only does math. The total set of acts it permits is enumerable.
The substrate is the ceiling. No actor, biological or silicon, no matter how sophisticated the selector, can make the substrate perform an act outside its permitted set. The substrate is also the floor. Every act incurs a thermodynamic cost — at minimum kBT · ln 2 per erased bit. A syscall emitted by a human and a syscall emitted by a daemon dissipate the same heat through the same gates.
Agency is what fits between the ceiling and the floor. The substrate confers it. The actor exercises it. The substrate is neutral to selection.
The shape of an act
Four axes locate every act.
Two are agent-native — who the actor is, and how the selection was made. Two are inherited — the primitive is a coordinate in compute, and the timestamp is universal. Every act in the system, by any agent, carries a value on each axis.
A
Actor
01 · WHO OCCUPIES THE SEATthe occupant exercising conferred agency
agent-native
σ
Selection process
02 · HOW THE ACT IS CHOSENdeterministic or stochastic
agent-native
p
Primitive
03 · WHICH SUBSTRATE OPERATIONa coordinate in compute.openie.dev
inherited from compute
t
Timestamp
04 · WHEN ON THE SUBSTRATE’S CLOCKuniversal — the act occurred at t
universal
The classification space
The 2 × 2 of agent configurations.
The classification of agents reduces to two values: who occupies the seat, and how they select. Every actor known to the framework lives in one of these four cells at any given moment, and many switch cells across acts.
⟨ H, S ⟩ source of novelty
Deliberation
Wetware noise drives selection. Judgment, taste, design, prose. The selection cannot be reproduced — not even by the same agent under the same conditions.
⟨ S, D ⟩ durable artifact
Compile, don’t loop
A compiler, a code generator, a scheduler, a language-model runtime in author mode. One stochastic upstream selection produces a deterministic program. The program is then a software-daemon agent in its own right.
⟨ S, S ⟩ no persistent artifact
Sample per iteration
A runtime samples each next act from a learned distribution. Every iteration re-samples. No persistent artifact survives the iteration. Tool-use loops, reinforcement-learning policies at inference.
The bill
Same task. Four cells. Four bills.
Extract invoice numbers from 500 PDFs. The substrate accepts all four routes. The cumulative cost across batches diverges by orders of magnitude.
⟨H, D⟩
$37.50
flat forever
⟨S, D⟩
$0.02
flat forever
⟨H, S⟩
$208
per batch · climbs
⟨S, S⟩
$7.50
per batch · climbs
The first two cells charge once. The second two charge every time. After ten batches, ⟨S, D⟩ is 3,750× lower-cost than ⟨S, S⟩, and the gap widens with every batch.
The seat has a history
“Computer” was a job title
before it was a machine.
Software daemons have occupied the agent seat alongside humans since the early Unix era. init is from 1971. cron from 1975. make from 1976. Every Unix system has had non-human agents selecting acts on its behalf for over fifty years.
The arrival of language-model runtimes is the addition of a new occupant to a seat with a long roster, not the invention of the seat.
What this site does not classify
Agency inherits.
This site classifies neither the substrate nor the agency itself — both are already classified by the family. It classifies the relationship: how acts are selected, by what kinds of occupants, at what cost, with what reproducibility, producing what durable artifacts.
compute the permitted-act set. The substrate enumerates the universe of acts an agent can select from. synthesis the per-act cost. Landauer floor and impedance mismatch fix the joules-per-act regardless of selector. state the shape of orchestration. Plans, machines, and transitions over acts. memory the substrate beneath each act. What each act writes to or reads from, and at what retention. communication the channel each act crosses. Bandwidth, latency, ordering, delivery guarantees. data the structures each act composes. Tables, indexes, columns, partitions. knowledge the object each act handles. Claims with seven axes attached: time, granularity, scope, certainty, provenance.
Companion stacks
One method, applied across fields.
Seven sibling sites plus the parent thesis. Each describes what systems do or what claims are. Agent describes what they all have in common: an actor selecting acts on the substrate, paying its bill in joules.
Compute
258 primitives, 33 families.
compute.openie.dev →
Synthesis
The seven-axis synthesizer stack.
synthesis.openie.dev →
Data
The ten-axis data-system stack.
data.openie.dev →
State
The nine-axis state stack.
state.openie.dev →
Communication
The eleven-axis channel stack.
communication.openie.dev →
Memory
The eleven-axis retention stack.
memory.openie.dev →
Knowledge
Seven axes · the object the others handle.
knowledge.openie.dev →
Stack
The overview · three names, one substrate.
stack.openie.dev →
Verify
Seven axes · joules per verified output.
proof.openie.dev →
OpenIE
The parent thesis.
openie.dev →