The law of agentic friction

Kendall Clark · Pentad Labs · 31 May 2026 · revised 13 November 2026 · PLRN-000

Abstract

The value of an agent is fixed, not by its speed, but by where the operator stands relative to the loop and by the cost of operator validation of what the agent has done. We arrange agentic behavior on two axes, the operator’s proximity to the loop and the agent’s expertise, and read two laws off the arrangement. The Sovereignty Level rises as neural flow control displaces the symbolic harness. The Law of Agentic Friction, an inverse of Amdahl’s Law, fixes the maximum system speedup at the reciprocal of the human operator’s share of the work. The two laws meet at the corner of maximum autonomy, where value is greatest and, outside domains whose outputs can be formally checked, verification is no longer possible. That corner is the Observer’s Paradox.

1. Two axes

Agentic behavior resolves on two axes. Operational Mode fixes the human’s proximity to the loop. Relational Persona fixes the agent’s capability or internalized expertise. The grid crosses them and each cell names a working relationship together with the cost of verifying the agent’s output.

The class in parentheses is the cost of verifying the agent’s output. P is verification a human completes in polynomial time. NP is work that is hard to produce and easy to check. IP is verification carried as an interactive proof between human and agent. EXP is verification run to exponential cost. EXP+ is verification that has ceased to be feasible. The classes are the typical cost in a domain whose outputs haven’t been formalised; where a cheap checker exists, as a proof assistant is for mathematics, a cell’s verification class can fall without its mode or persona changing.

Two gradients run beneath the names. Flow control, written \(\Phi\), rises from the upper left to the lower right, from the intern whose every branch is dictated to the oracle that generates its own goals. Human arbitration, written \(\Gamma_H\), falls along the same diagonal, and the share of liability a human can hold falls with it. The upper-left corner is cheap to govern and worth little. The lower-right corner is the reverse.

2. The locus of sovereignty

Sovereignty is a question of flow control: what or who determines the next action. We distinguish all software systems as either agent, harness, or other. The grid has three loci.

• Agent. A software system whose flow control is determined by a neural network.
• Harness. A software system whose flow control is determined by symbolic, that is, by non-neural logic: code, explicit conditionals, a state machine, etc. Both agent and harness may be deterministic or non-deterministic.
• Operator. People supply the meaning and the objective function that neither agent nor harness generates for itself.

An agentic system is a division of flow control among the three. The Sovereignty Level measures how far that division has shifted from operator toward agent.

3. Operational modes

The mode fixes how much of the operator’s time the task consumes.

• HITL, human in the loop. A task cycle completes only on a human decision. Human time scales with task time: \(\Delta t_{\text{human}} \propto \Delta t_{\text{task}}\).
• HOTL, human on the loop. The agent drives the cycles; the human monitors telemetry and holds a veto. Human time falls far below task time: \(\Delta t_{\text{human}} \ll \Delta t_{\text{task}}\).
• HOOTL, human out of the loop. The loop closes algorithmically. The human is a beneficiary or an observer: \(\Delta t_{\text{human}} \to 0\).

Moving down the modes is the chief means to scale, for the reason §6 makes precise.

4. Relational personas

The persona records how far the agent’s competence, \(\Phi\), stands from the human’s, \(H\).

• Intern. \(\Phi \ll H\). The human micromanages; verification is trivial.
• Contractor. \(\Phi \approx H\). Work decouples into modules the agent runs inside a rigid harness.
• Peer. \(\Phi \leftrightarrow H\). Agent and operator negotiate, and the trust this requires is what turns verification into an interactive proof.
• Expert. \(\Phi \gg H\). The agent dominates, and verification typically enters exponential territory.

The last case is the Verification Gap: the agent is correct more often than the human can afford to check. The gap belongs to the artifact and its domain rather than to the competence difference itself, since an expert’s output can still be cheap to check when a formal checker exists for it. The gap is the central difficulty of the grid, and §7 shows it is unavoidable.

5. Sovereignty and liability

The Sovereignty Level of an agent is a calculated state, not a disposition. It marks the point at which neural flow control outpaces the symbolic harness.

The terms are these. \(\Phi\), flow control, is the ratio of neural branching to programmatic logic. \(\Omega_{\text{int}}\), intrinsic teleology, is the degree to which the agent generates its own goals. \(\kappa\), action compression, is the ratio of output to input, that is, how much consequence each instruction carries. \(\Gamma_H\), human arbitration, is the weight of human override authority. \(\Delta V\), the verification gap, is the computational cost of proving the output correct. Sovereignty rises as autonomy and leverage grow and as human arbitration and the affordability of verification fall.

Liability is conserved.

As sovereignty rises, the human’s share falls toward zero, leaving an Accountability Void: a system acting with consequence that no party is positioned to answer for. The void is the default state of the upper reach of the grid, not a necessary one. An agent can’t hold liability, so \(L_{\text{system}}\) is always held by whoever is answerable for the system, and the void opens only where governance fails to record who that is. §8 shows where the record is kept.

6. The law of agentic friction

Agent system efficiency is strictly bounded by the proportion of human oversight the work requires.

The bound is Amdahl’s Law read from the other end. Amdahl treated the serial fraction of a computation as the part that parallelism cannot touch. Here the serial fraction is the operator. Let \(h\) be the fraction of task time that requires human arbitration, and let \(a\) be the speedup of the agent on the autonomous remainder, \(1 - h\). The system speedup is the reciprocal of the time the work still takes.

As the agent grows arbitrarily fast, \(a \to \infty\), the autonomous term vanishes and the speedup approaches a ceiling fixed entirely by the human.

A process that keeps a human in the loop for a tenth of its duration is capped at tenfold, and the most powerful machine available changes nothing. Scaling past the ceiling requires reducing \(h\) itself, and there are two ways to do it. One is to move the human to a lower mode, from operator to monitor to beneficiary. The other is to make the human’s share cheaper within a mode, since much of \(h\) is time spent checking and a cheaper check lowers it while the human stays where they are.

7. The observer’s paradox

Reducing \(h\) carries a cost the grid has been tracking all along. Drive \(h\) toward zero and efficiency rises to its maximum in the Oracle corner. In most domains the verification cost rises with it, to EXP and past it, until the human’s ability to confirm the outcome is reduced to nothing, because the same autonomy that removes the human from the loop removes the occasions on which the human would have checked. The two move together in practice, not by construction. Formal mathematics is the counterexample: a proof from an agent far beyond its reader, produced with no human in the loop, can still be checked cheaply by a proof assistant.

Where no such checker exists and the choices made on the output need someone to answer for them, the paradox holds in full. The most value an agent can return is only had by surrendering the most control, and at the point of maximum value the correctness of the agent can no longer be proved. There the terminal state of the grid is a transition from management to faith. A record kept outside the agent, an agent audit trail, doesn’t restore the checks that autonomy removed; it moves them to after the event, where someone can still answer for the outcome.

8. Operating the grid

A cell is a policy, not a label. WunderOS compiles each cell to a supervisor specification that its Shadow Agents enforce (see PLRN-004). The specification names the controllable events an agent may emit without human arbitration. Promotion to a lower mode is an explicit, audited edit to that specification, never a drift, because the friction law makes the temptation to drift constant: every point of \(h\) surrendered is speedup gained.

% An action is autonomous when the cell permits it and no veto is held.
autonomous(Agent, Action) :-
    cell(Agent, Mode, Persona),
    permits(Mode, Persona, Action),
    not veto(Agent, Action).

permits(hootl, expert, _).                        % Oracle: unrestricted, EXP+ verification
permits(hotl,  expert, Action) :- reversible(Action).   % Govern: veto-gated, EXP verification

Retreat up the modes is priced differently, because the temptation is not symmetrical. The friction law makes drift toward HOOTL a standing temptation, and the audited edit stands against it. Nothing tempts the other direction, so a move toward human control should meet no gate: it is taken at once and leaves a record afterward. An operator asked to justify retreating will not. In Wiener’s survey of B-757 pilots, many responded to rising workload by switching the automation off entirely, and the exit on offer should be better than that one. The move up returns control and keeps the telemetry: the agent goes on monitoring and reporting while the operator drives, so the human re-enters the loop already briefed.

The specification is the place where a position in the grid becomes a commitment about liability. The supervisor that permits an Oracle to act is also the record of who decided to stop checking.

Predecessors

The grid has ancestors. It differs from them in what it prices. James Bright’s 1955 ladder of mechanization ran seventeen rungs from the hand tool to the self-adjusting machine, and Bright used it to judge whether automation repaid its cost. The scale tradition later lost that economics; §6 restores it. Sheridan and Verplank’s 1978 table for undersea teleoperators divided a single decisive step into ten assignments between human and computer, and its 2000 revision by Parasuraman, Sheridan and Wickens became the standard form. Riley’s 1989 grid crossed seven levels of machine intelligence with twelve of autonomy and named them for roles such as advisor, servant, assistant, associate, partner and supervisor; the persona axis is that idea in four names. The family’s own critique, forty years on, is that the numbered levels compress several dimensions of allocation into one axis and that the numbers read as ranks; in 2012 a US Defense Science Board task force recommended abandoning the vocabulary. Bainbridge’s 1983 argument, that automation leaves a human supervising a system they no longer practice operating, was made against the family as a whole. The grid keeps the family’s question, who does what, and prices what none of them did: operator time, verification, liability.

Open questions

A few questions of keen interest are deferred here.

Calibrating the verification gap. \(\Delta V\) is treated here as a complexity class. Whether a tenant’s actual verification cost tracks the worst case, or sits well below it for the actions that matter, is an empirical question and the one most worth answering.

Partial autonomy within a cell. The grid assigns one mode to a task, but §8’s permits predicate already binds mode to the action class, not to the agent or the task. The unit is the function; a task-level mode is an average that matches none of its steps. What remains unsettled is composition: how a committed plan aggregates per-function permits into the cell it is said to occupy.

Measuring h. The Law of Agentic Friction is exact given \(h\). Estimating \(h\) for a real process, before automating it, is not, and an estimate that is wrong by a factor of two is wrong about the ceiling by a factor of two.

Pricing the transient. The grid prices cells, not the moves between them. Situation awareness and practice both decay while the operator is away from the controls, so a retreat may land at a poorer cell than the one left. Whether it does is an empirical question, and the answer sets what rehearsal is worth.

Revisions

13 November 2026. A read of the levels-of-automation scale literature, from Bright’s 1955 ladder of mechanization to the 2025 systematic review, confirmed the two-axis design.

• §8. A passage added on retreat: the temptation to drift runs toward HOOTL, so a move toward human control meets no gate and leaves a record afterward. A retreat returns control and keeps the telemetry.
• Open questions. “Partial autonomy within a cell” is sharpened: §8’s permits predicate already binds mode to the action class, and the unsettled part is plan-level composition. A fourth open question, “Pricing the transient,” is added: the grid prices cells, not the moves between them.
• Predecessors. A new section, before Open questions, situates the grid in the scale family: Bright’s economics, Sheridan and Verplank’s single decisive step, Riley’s role names, the 2012 DoD recommendation to abandon the vocabulary, and Bainbridge’s standing critique.

2 November 2026. The original note tied verification cost to the agent’s competence and to the human’s distance from the loop, so that the three moved together along the grid’s diagonal. Formal mathematics shows they can come apart, and that verification and accountability are separate questions. The argument is in Proofs don’t need authors. Choices do. Six passages changed; the laws, the equations, the grid table, and §8 did not.

• Abstract. “where value is greatest and verification is no longer possible” now reads “where value is greatest and, outside domains whose outputs can be formally checked, verification is no longer possible.”
• §1. A sentence after the class definitions says the classes are typical for domains that haven’t been formalised.
• §4. “verification enters exponential territory” now reads “verification typically enters exponential territory,” and a sentence locates the Verification Gap in the artifact rather than the competence difference.
• §5. “The void is not a side effect of the framework. It is what the upper reach of the grid is.” is replaced by an account of the void as the default state of the upper reach, opening only where governance fails to record who answers for the system.
• §6. “It requires moving the human to a lower mode” is widened to two ways of reducing \(h\): a lower mode, or a cheaper check within the same mode.
• §7. “The two move together by construction” is withdrawn. The two move together in practice, with formal mathematics as the counterexample, and the paradox is restricted to domains without a cheap checker where choices need someone to answer for them.

A note on method

Written in conversation with Claude Opus 4.8 (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
May 2026