Productivity Voids: Agentic Economics at Firm Scale

Kendall Clark · Pentad Labs · 12 August 2026 · PLRN-024

Abstract

An enterprise puts agents on every desk. Adoption reaches nearly everyone, and the individual gains are real and large. Work that took an afternoon now takes three minutes, and the people doing it will say so. At the end of the year the firm produces roughly what it produced before. Headcount is the same, revenue is the same, and the same work ships on the same dates.

This is not a failure of adoption, of tech or tooling, or of will. It is what the math predicts, and the math isn’t difficult.

PLRN-000 arranges agentic work on two axes: how close the human stands to the loop and how far the agent’s competence stands from the human’s. Then it reads two laws off the arrangement. The Law of Agentic Friction fixes the maximum speedup at the reciprocal of the human’s share of the work. A process that needs a person for a tenth of its duration cannot go more than ten times faster, no matter how fast the agent, LLMs, or GPUs become.

The Conservation of Liability fixes the total consequence at one and lets only its shares move. As autonomy rises, the share a human can answer for falls toward zero, and PLRN-000 calls what is left behind the Accountability Void.

Both laws are derived for one person working with one agent. In that case the operator supplies the objective function, so the operator’s interest and the system’s interest are the same thing by construction.

But a firm is not that. It is a population of such cells with a principal above them, and the party who pays for the work is not the party who decides how it gets done.

Under that split the friction law survives and stops composing, because a cell’s speedup reaches the firm only when the operator’s step has no float. The conservation of liability does not survive at all. The surplus an agent creates has three destinations, and nothing measured at the operator tells them apart. So the quantity both laws depend on stops being a measurement and becomes a report from a party with an interest in the answer.

What a firm can honestly measure follows, and so, too, does what it cannot.

1. The simplification PLRN-000 made

Our grid in PLRN-000 names three loci:

  1. the agent, whose flow control comes from a neural network,
  2. the harness, whose flow control comes from (for simplicity here) symbolic logic, and
  3. the operator, who supplies the meaning and the objective function that neither of the others generates.

Nothing in the framework asks whose objective function it is. With one human there is only one answer. That is not a defect. It is what makes the single-cell case tractable, and it is true of a person working with a tool. The person who owns the goal, holds the keyboard, carries the consequence, and collects the benefit is one person. Their interest and the system’s interest coincide because they and the system are the only things in this model and they wrote the interest down.

A firm is a different object, and the difference is not one of scale only.

2. In the real economy, there is a fourth locus

In a firm the operator role divides.

  1. One party supplies the objective function, pays for the work, and answers for it in law. Call that party the Principal.
  2. Another party sits at the keyboard, chooses how the work is done, and receives the benefit of doing it faster. Call that party the Operator.

PLRN-000’s single locus is these two collapsed, and they come apart the moment the person who benefits from a decision is not the person who bears its cost.

Both parties want the usual things. The Principal wants more output for each dollar. The Operator wants more dollars, or more hours of slack, for each unit of arbitration. Neither want is a failure of character, and treating either as one will produce a design that does not work. A reason that belongs to an institution moves a person only where it meets a motive that person already holds. The firm’s reason to want throughput meets nothing in the Operator unless the firm arranges for it to, through profit-sharing, for example.

So the grid acquires a fourth locus, and the two laws must be read again with four.

3. Float, and why the friction law stops composing

The Law of Agentic Friction fixes the maximum speedup at \(1/h\), where \(h\) is the fraction of task time that requires human arbitration. Read at firm scale the law invites an error, which is to suppose that a firm whose operators have driven their \(h\) toward zero has thereby raised its own ceiling. It has not. The reason is visible with a clock.

A managing director sends a presentation back at 23:00 on a Tuesday. The client meeting is at 09:00 Wednesday. The process window is ten hours and the analyst’s work is six, so the task carries four hours of float: its duration can grow by four hours, or shrink by any amount, and the meeting still happens at 09:00.

Before the agents, the analyst works until 05:00 and sleeps three hours. After the agents, the revisions are done at 23:04 and the analyst sleeps eight hours and spends one on Netflix.

The process took ten hours in both worlds. The firm’s speedup is 1.0.

The analyst’s \(h\) fell by more than ninety-nine percent and the firm gained nothing, because the deck was never the thing that decided when the process ended. The calendar decided.

But change one number and the result inverts. Now a risk model must rerun before the 09:30 market open. The run takes seven hours inside a seven-and-a-half hour window, so float is thirty minutes. Cut that run to four minutes and the firm can rerun after a late data revision, run three scenarios where it ran one, or move the deadline in. That is a new capability the firm did not have.

The cell speedup is of the same order in both cases. The firm gains nothing in the first and gains a new capability in the second, and the difference is float, which is a property of the schedule rather than of the worker.

The firm therefore has its own serial fraction, written here as \(H\), the share of process duration that does not shrink when agent-assisted steps get faster. It is made of the client meeting, the reviewer’s availability, the compliance window, the approval chain, and the counterparty’s response time.

Amdahl’s Law applies again at this level and gives the firm its ceiling.

\[S_{\text{firm},\max} = \frac{1}{H}\]

\(H\) is not built from the operators’ values of \(h\). The two quantities run over different denominators, task time in one case and process duration in the other, and an operator whose step sits in float can reach \(h = 0\) while contributing nothing to \(H\). The friction law holds exactly at both levels and does not compose between them.

Reducing \(h\) is one project. Reducing \(H\) is a different project, and it is a redesign of the process, which likely hits non-negotiable external limits (market regulations and opening times, for example), rather than an adoption of a tool.

4. Three destinations, one measurement

An agent that removes six hours of human arbitration from a task creates six hours of something. That something has three destinations.

  1. The firm converts it to output, which happens when the step was on the binding constraint.
  2. The Operator takes it as slack, which happens when the step was not on the binding constraint and the Operator is free to leave.
  3. It evaporates, which happens when the step was not on the binding constraint and the Operator waits at the desk until the deadline that never moved arrives anyway.

The third destination is a Productivity Void: an hour of arbitration removed from the process, paid for by the firm, and received by no one.

The three have different consequences for the firm and they produce an identical measurement. In all three cases the same task completed, the same arbitration time disappeared, and \(h\) fell by the same amount. Nothing an observer can read off the Operator’s record distinguishes them, because what distinguishes them is not in the Operator’s record. It is in the shape of the process around the Operator, and it is exactly the quantity that no one instrumented.

This is why an enterprise can report that most of its employees use agents daily and report no change in what it produces. Both statements can be true at once, and the framework predicts the pair rather than treating it as a puzzle.

5. Liability does not survive the split

PLRN-000 states that liability is conserved.

\[L_{\text{human}} + L_{\text{system}} + L_{\text{residual}} = 1\]

The law is exact for one cell. At firm scale it fails, and it fails on a term rather than on the math. \(L_{\text{human}}\) means the Operator, because in the single-cell case there is no other human it can mean. Under the split it must divide into an Operator share and a Principal share, and those two are fixed by different mechanisms. The Operator’s share is fixed by the cell, which is to say by how much arbitration authority the position leaves in human hands. The Principal’s share is fixed by law and by contract, and it does not move when the cell moves. A firm answers for what its agents did whether or not the Operator declared the mode under which they did it.

Two quantities set by non-commensurable mechanisms do not sum to a conserved total. So the conservation law holds within a cell and breaks across the firm, and what it breaks into is worse than an imbalance.

PLRN-000 names the Accountability Void, in which sovereignty rises until no party is positioned to answer for a system acting with consequence. The firm scale produces a different void earlier along the same diagonal. Here a party is positioned to answer, and it is not the party that decided.

The Principal holds the liability for a cell the Principal never chose, because the choice was made at a keyboard and never declared. Call this the Delegation Void.

The Delegation Void is sharper than the Accountability Void, because it does not require the far corner of the grid. It appears as soon as an Operator can move down a mode without saying so, which is to say immediately.

6. The message and the record

PLRN-000 defers the measurement of \(h\) as an empirical question and observes that an estimate wrong by a factor of two is wrong about the ceiling by a factor of two. That is the single-cell difficulty and it is real. At firm scale a second difficulty sits on top of it and is worse.

In one cell, \(h\) is observed. The Operator is the Principal and watches their own clock, so the quantity has no channel to travel through and nothing to be distorted by. In a firm the Principal learns \(h\) from the Operator. The quantity is no longer measured, it is reported, and it is reported by the party whose position depends on the answer. An Operator who discloses that six hours became sixty seconds invites one of two responses, a smaller analyst class or five times the work, and neither is a response that the Operator wants, quite understandably. Concealment is the rational move and it does not require bad faith, only the ordinary reading of one’s own situation.

\(h\) therefore stops being a measurement and becomes an inevitably-contested message, that is, there is a fact of the matter, and the dispute belongs to its reporting. The state is guaranteed wherever four conditions hold at once:

  1. the quantity has a determinate value,
  2. one party alone is positioned to observe it,
  3. that party’s material condition or standing moves with the answer, and
  4. the party who needs the answer cannot check it independently.

Given those four, disagreement follows without anyone acting in bad faith. It is not permanent, though, because it ends when the second condition fails.

From the firm’s perspective, what level and sorts of operator observation, if any at all, are consistent with the amount and quality of operator arbitration the firm’s competitive operation requires is the question §7 takes up.

What converts a message back into a record is the supervised path. PLRN-000 compiles each cell to a supervisor specification which WunderOS Shadow Agents enforce, and the specification names the controllable events an agent can emit without human arbitration. Promotion to a lower mode is an explicit and audited edit to that specification. The mechanism was designed to answer a liability question, namely who decided to stop checking, and at firm scale it answers a second question it was not built for: what the Operator’s cell actually was, as opposed to what the Operator said it was.

The property that makes this work is coverage. The supervisor sees the actions that route through it, and an action taken by another route leaves no record and therefore does not exist for any purpose, liability included. So a firm’s ability to hold a record of \(h\) is exactly its ability to make the supervised path the only path. That is a general property of any WunderOS deployment and not a feature of one.

WunderOS’s first enterprise application, the Agentic Data Enclave, is the limit case rather than the frame. A boundary that data cannot cross makes the supervised path the only path by construction, so the enclave holds the strongest form of a property that weaker deployments hold in weaker forms. A firm without one holds the same record by policy and by network control, which is to say less completely, and §8 says how much less.

7. Measuring the Operator directly, and what it costs

A firm that needs \(h\), and cannot get a neutral report of it from the Operator, has two routes to it.

This is where capital and labor converge, the second bowtie of the agentic era, and the reading here is from the firm’s side. The second route is surveillance. Whether it pays turns on a single variable, which is the kind of arbitration the firm’s operations require.

Arbitration divides by discretion. Compliance arbitration applies a known rule to a case, so the rule determines the answer, the answer is independently checkable, and the Operator’s contribution is legible in the artifact. Judgment arbitration decides where no rule determines the answer, and it needs the Operator to notice a problem nobody asked about, to say so, and to own the call.

Where operations need compliance arbitration, surveillance is cheap and close to harmless. The Operator has little discretion to withhold, so there is little to suppress, and the work product already checks the report. Where liability attaches to timeliness rather than to judgment, as in processing work against a service level, measuring time measures the thing that matters. A firm in that position can instrument its Operators and lose nothing it wanted.

Where operations need judgment arbitration, the account reverses, in four ways.

First, h-measurement taxes the three things judgment arbitration is made of. Noticing a problem nobody asked about takes unbilled time, which reads as idleness. Raising the flag creates work and delay, which reads as slowness. Owning a call requires believing that no number will be produced later to second-guess it. The instrument penalizes each part, and a flag never raised leaves nothing to measure, so the loss is invisible to the instrument that caused it.

Second, h-measurement raises the personal cost of routing through the supervised path. The rational response is not to stop using agents but to use them where the supervisor looks least. Consequential action mostly cannot leave that path, since it needs credentials and systems the firm controls. Deliberation leaves easily. The Operator decides off the path, executes on it, and declares a cell in which a human held the veto, when in fact an unrecorded agent decided and the human forwarded the result. The record does not acquire a gap. It acquires a false entry, and it acquires one exactly for the Operators with the most reason to produce one. A liability record with acknowledged gaps is usable. A liability record that confidently misstates who stopped checking is worse than none, since the sentence it exists to support is the one it now gets wrong.

Third, and this one falls out of the grid’s own geometry, surveillance increases the quantity it was bought to decrease. A measured Operator seeks cover, and seeking cover means getting a human decision on record before acting. That is HITL by definition. The working relationship moves toward the upper left of the grid, which PLRN-000 calls cheap to govern and worth little, and where human time scales with task time. The firm instruments its Operators to drive \(h\) down and the instrument drives \(h\) up. The effect compounds, because a higher \(h\) reads on the instrument as worse performance, which invites closer measurement.

Fourth, and plainest, perfect measurement still answers the wrong question. Suppose the surveillance works and returns the Operator’s true \(h\). §3 and §4 have already shown that \(h\) alone cannot say whether anything was bought, since the same measurement is consistent with output, with slack, and with a Productivity Void. A firm can therefore spend real money and real trust to obtain an accurate number that does not answer the question which motivated the spending.

The four together give the shape of it. Surveillance is least harmful where it matters least and most harmful where the firm most wants the answer, because the value of the arbitration and the damage done by measuring it rise together.

The firm still wants the number, reasonably so. What the number needs is a second one beside it. Hours saved are intermediate and not final. In the agentic economics subsystems WunderOS is building at publication time, a reduction in active human labor time is credited as value only when it is paired with an observable downstream signal: more work completed, faster revenue, lower contractor spend, avoided hiring, reduced backlog, or redeployed capacity. It is never credited when the system reports the saving on its own authority. That rule governs the reward signal WunderOS optimizes against internally, which is the only real test of a measurement discipline, namely whether the party stating it pays for it.

None of this expects the Principal to want less. It observes that the two things the Principal wants are inseparable. The good on offer is authorized agentic action at reduced \(h\), and the restructuring of accountability that §5 describes is constitutive of that good rather than an accessory to it. A firm that takes agent-speed work and keeps pre-agent accountability has not bought a cheaper version of the good. It has bought the Delegation Void.

No one avoids the unavoidable consequences of the goods they choose, and a vendor who offers to arrange otherwise is selling the arrangement, not the good.

8. Where the argument is weaker

The off-path argument varies in strength with the work. Where the deliverable is, say, text a person can retype, an unobserved channel is always available and the false-cell mechanism is strong. Where the work must touch a system of record through controlled credentials, the action stays on the path and only the deliberation escapes, so the mechanism weakens to a claim about the cell label rather than about coverage. It weakens further where the firm’s liability attaches to the action alone and does not reach the basis for the decision. It does not weaken to nothing anywhere, because the cell label is what the record exists to certify.

The reading of \(h\) and \(\Delta V\) as two quantities is PLRN-000’s own and is retained here. Arbitration cost and verification cost are separate terms in the grid and they do not collapse. A firm that reduces the first while raising the second has moved labor rather than removed it, and the net is what the downstream signal has to show.

Open questions

Measuring \(H\). The firm’s serial fraction has the same difficulty as \(h\) and one more. A process is spread across parties who each see one segment of it, so no participant is positioned to report the whole, and the binding constraint is often held by someone outside the firm. Whether \(H\) can be instrumented at all without a process model the firm does not currently maintain is an open question.

Whether the downstream signal exists at that resolution. The signal is stated as a condition on a per-operator claim, and it is not obvious that any firm can supply one at per-operator resolution. Revenue and backlog are firm-level or team-level quantities. If the signal cannot descend to the resolution the claim is made at, then per-operator value is not merely hard to measure but unavailable, and saying so plainly is better than leaving a condition standing that nothing can satisfy.

Whether float is stable. Instrumenting the binding constraint assumes it holds still long enough to be measured. In processes where it moves between steps and between weeks, a measurement of float is a statement about the past, and the value of a per-operator claim decays at whatever rate the constraint migrates.

A note on method

Written in conversation with Claude Opus 5 (Anthropic) as structured interlocutor and prose editor. The framework, the claims, and the architectural commitments are mine.

Kendall Clark · k@pentad.ai
—Great Falls, Virginia
August 2026