Total Cost of Agency (TCA) Explained

• Total Cost of Agency (TCA) refers to the aggregate resources needed for senses, acquiring information, forming decisions, coordinating with systems, and producing actionable outcomes. It also encompasses physical, computational, informational, operational, cognitive, contractual, and governance resources necessary for agency.
• TCA is applied in numerous fields such as thermodynamics, decision theory, finance, and AI deployment. For instance, in thermodynamics, TCA involves the work and entropy production; in finance, it measures implicit execution costs like spread and price impact; in AI, it includes costs associated with inference, infrastructure, oversight, and debugging.
• Understanding TCA helps in optimizing agency by identifying and minimizing the costs associated with each component, such as decision-making, actuation, and governance, to enhance efficiency, reliability, and performance within various systems.

Total Cost of Agency (TCA) is a cross-domain term for the aggregate resources consumed when an agent or delegated system senses, acquires information, forms or implements decisions, coordinates with other actors or systems, and produces actions or outcomes. The term has no single universally accepted definition across the cited literature. It denotes a thermodynamic cost of information-enabled control, an expected accumulated cost in Markov decision processes, a financial and organizational cost of delegated execution, or a broader lifecycle cost of human–AI and institutional agency. The common analytical problem is to measure not only the benefit produced by agency, but also the physical, computational, informational, operational, cognitive, contractual, and governance resources required to make that agency possible.

1. Scope and conceptual boundaries

TCA is best treated as a family of related constructs rather than a settled metric. The broadest formulation decomposes agency into sensing, decision or control, actuation, and dissipation, with benefits accounted for separately:

$$\text{Total cost of agency} = \text{sensing cost} + \text{decision/control cost} + \text{actuation cost} + \text{dissipation} - \text{benefit}.$$

This formulation is an interpretive decomposition rather than a universal identity. Different domains assign different meanings to “cost,” “agency,” and “benefit.”

A thermodynamic interpretation treats agency as physical information processing. In a minimal information engine, measurement creates mutual information that enables work extraction, but the measurement operation itself requires energy. The resulting inequality is

$$\beta_R\langle W_{ex}\rangle \leq \Delta I_M \leq \beta_M\langle W_{sup}\rangle,$$

where $\Delta I_M$ is measurement-generated mutual information, $\langle W_{ex}\rangle$ is extracted work, and $\langle W_{sup}\rangle$ is measurement-supply work. The result establishes that information-enabled benefit cannot be separated from the physical cost of acquiring actionable information (Um et al., 2015).

In decision theory, TCA can instead be represented by the expected accumulated one-stage cost of a controlled Markov process. The term is not defined as a distinct economic concept in the relevant Markov decision process literature; it becomes a TCA model only when agency-related costs are included in the one-stage cost function $c(x,a)$ and the state captures all economically relevant information (Feinberg et al., 2017).

In market execution, transaction cost analysis (TCA) measures implicit execution costs, including spread and price impact. This usage is narrower and financially specialized, but it contributes measurable components to a broader agency-cost account: the cost of delegating execution to brokers, algorithms, venues, and counterparties (Guo et al., 2019, Markov, 2019).

In generative-AI systems, TCA concerns the cost of producing reliable delegated action rather than merely generating model outputs. Relevant components include inference, infrastructure, memory injection, human oversight, debugging, rework, governance, and loss of autonomous capacity. The term is explicitly formalized for multi-agent LLM workflows as a decomposition of base-prompt, inference, memory-injection, miss-penalty, and context-accumulation costs (Singh et al., 20 Sep 2026).

Distinct uses of TCA

Domain	Primary object	Principal cost basis
Thermodynamic agency	Information-enabled physical control	Work, heat, and entropy production
MDPs	Sequential agency decisions	Expected total or average one-stage cost
Finance	Delegated trade execution	Spread, impact, timing, and reversion
Human–AI interaction	User control over AI-mediated action	Cognitive, interactional, and performance costs
AI deployment	Computationally delegated output or action	Lifecycle, inference, oversight, and failure costs
Public administration	Distributed operation of a tax system	Government and taxpayer compliance resources

The terms should not be conflated. Transaction cost analysis, total administrative cost, total cost of ownership, and TCA may overlap, but none is automatically equivalent to a comprehensive cost of agency.

2. Thermodynamic foundations of agency

The most direct microscopic account of TCA models a thermally relaxing two-level system controlled by a one-bit memory. The physical system has states $s\in\{0,1\}$ separated by energy $\Delta E$ and coupled to a heat bath. The controller has memory states $m\in\{0,1\}$. A cycle consists of relaxation, measurement, and feedback.

During relaxation, the system approaches equilibrium. During measurement, the memory copies the system state with error probability $\epsilon$, where $$\beta_R\langle W_{ex}\rangle \leq \Delta I_M \leq \beta_M\langle W_{sup}\rangle,$$0. During feedback, the system is flipped conditional on the memory state. Correctly identifying the excited state permits extraction of $$\beta_R\langle W_{ex}\rangle \leq \Delta I_M \leq \beta_M\langle W_{sup}\rangle,$$1; an erroneous record can instead consume the same amount of work.

The mutual information

$$\beta_R\langle W_{ex}\rangle \leq \Delta I_M \leq \beta_M\langle W_{sup}\rangle,$$2

is generated by measurement and subsequently consumed by feedback and relaxation. The information is therefore physically redistributed rather than merely “used” abstractly. In the stationary cycle,

$$\beta_R\langle W_{ex}\rangle \leq \Delta I_M \leq \beta_M\langle W_{sup}\rangle,$$3

The associated entropy changes satisfy

$$\beta_R\langle W_{ex}\rangle \leq \Delta I_M \leq \beta_M\langle W_{sup}\rangle,$$4

The feedback map is deterministic and reversible, so it preserves joint Shannon entropy and has zero direct entropy production in the idealized implementation. This does not imply that agency as a whole is costless: the sensing operation remains a physical stochastic process with its own entropy production.

The engine’s extracted work is

$$\beta_R\langle W_{ex}\rangle \leq \Delta I_M \leq \beta_M\langle W_{sup}\rangle,$$5

while the measurement apparatus receives supplied work $$\beta_R\langle W_{ex}\rangle \leq \Delta I_M \leq \beta_M\langle W_{sup}\rangle,$$6. At stationarity,

$$\beta_R\langle W_{ex}\rangle \leq \Delta I_M \leq \beta_M\langle W_{sup}\rangle,$$7

For equal reservoir temperatures, the net work of the complete apparatus cannot be positive. The measurement cost compensates the apparent benefit of feedback:

$$\beta_R\langle W_{ex}\rangle \leq \Delta I_M \leq \beta_M\langle W_{sup}\rangle,$$8

Thus, the thermodynamic TCA principle is an information–benefit–cost chain:

$$\beta_R\langle W_{ex}\rangle \leq \Delta I_M \leq \beta_M\langle W_{sup}\rangle,$$9

Finite-time operation introduces a cost–benefit trade-off. Short measurement times reduce accuracy and mutual information; short relaxation times fail to repopulate the excited state. Long cycles increase work and efficiency per cycle but reduce power because cycle duration increases. The model has a stall time $\Delta I_M$0 at which extracted and supplied work are equal, and an optimal-power time $\Delta I_M$1 satisfying $\Delta I_M$2.

The model also differs from conventional reversible heat engines. Maximum efficiency is approached as $\Delta I_M$3, but entropy production per cycle is then maximal, while entropy-production rate tends to zero. The long-cycle limit performs increasingly complete nonequilibrium relaxation and measurement; it is not a quasistatic reversible traversal of equilibrium states.

The thermodynamic model directly establishes three TCA-relevant facts:

1. Information acquisition has a physical minimum cost.
2. Information bounds control-enabled benefit.
3. The complete system-plus-controller-plus-environment obeys the conventional second law.

It does not provide a general theory of cognition, planning, memory hierarchy, autonomous learning, or multi-stage agency. The controller performs measurement and control simultaneously, no separate reset protocol is modeled, and instantaneous feedback suppresses finite-speed actuation costs.

3. Sequential decision and institutional agency

A second interpretation embeds TCA in an undiscounted Markov decision process. Let $\Delta I_M$4 be the state space, $\Delta I_M$5 the action space, $\Delta I_M$6 the feasible actions, and $\Delta I_M$7 the one-stage cost. A controlled transition kernel $\Delta I_M$8 may be substochastic or weighted, with total mass

$\Delta I_M$9

The expected total cost under policy $\langle W_{ex}\rangle$0 is

$\langle W_{ex}\rangle$1

and the optimal value is

$\langle W_{ex}\rangle$2

If the transition kernel is normalized, $\langle W_{ex}\rangle$3, and the criterion reduces to the expected accumulated one-stage cost. The total-cost optimality equation is

$\langle W_{ex}\rangle$4

For long-run agency operation, the average-cost criterion is

$\langle W_{ex}\rangle$5

Under the stated conditions, a constant average cost $\langle W_{ex}\rangle$6 and a bias function $\langle W_{ex}\rangle$7 satisfy

$\langle W_{ex}\rangle$8

Encoding agency costs

A TCA-oriented MDP can place operational costs in

$\langle W_{ex}\rangle$9

Hidden information, trust, contract status, relationship history, and accumulated interactions must be incorporated into the state when they affect future costs or transitions. Incentive choices, audits, reporting requirements, and delegation levels become action variables.

A direct TCA interpretation requires that costs be additive over periods, relevant information be representable by a Markov state, costs be nonnegative or shiftable to nonnegative values, transition kernels satisfy the required measurability and continuity conditions, and the model satisfy a transience or hitting-time condition.

The reduction developed for total-cost MDPs transforms the problem into a discounted MDP by adding an absorbing state, constructing a superharmonic function $\langle W_{sup}\rangle$0, and defining transformed costs and transition probabilities. The resulting value relation is

$\langle W_{sup}\rangle$1

Under weak continuity, compact-valued continuous action sets, lower semicontinuous costs, and a uniform transience condition, stationary deterministic optimal policies exist. For average-cost problems, a reference state $\langle W_{sup}\rangle$2 and a uniform hitting-time condition support a related Hoffman–Veinott–Altman–González transformation. The transformed discounted value generates the average-cost pair

$\langle W_{sup}\rangle$3

These results establish a computational route for TCA when agency is modeled as a controlled sequential process. They do not constitute a theory of strategic agency. Principal–agent interaction, equilibrium behavior, partial observability, and multiple optimizing actors may require stochastic games, constrained MDPs, or partially observed MDPs.

An inventory-control example illustrates the framework. Ordering, holding, lost sales, and capacity constraints form a controlled Markov process. The model can optimize expected pre-loss total cost or long-run average cost through a transformed discounted problem. The example demonstrates that “total cost” is mathematically an accumulated operational criterion; the agency interpretation must be supplied through the state, action, cost, and transition specifications.

4. Contractual, administrative, and execution costs

Principal–agent contracts

In Bayesian agency, a principal chooses an outcome-contingent contract for an agent whose type and action are unobserved. Each type $\langle W_{sup}\rangle$4 has action cost $\langle W_{sup}\rangle$5 and outcome distribution $\langle W_{sup}\rangle$6. A contract $\langle W_{sup}\rangle$7 generates expected compensation

$\langle W_{sup}\rangle$8

The agent’s utility is

$\langle W_{sup}\rangle$9

and the principal’s utility is

$c(x,a)$0

The identity

$c(x,a)$1

separates total surplus from its distribution between principal and agent.

A TCA interpretation can decompose the first-best gap into action inefficiency and agent rent:

$c(x,a)$2

This decomposition is conditional on a specified first-best benchmark. The paper does not define a universal TCA measure, and compensation is partly a transfer rather than necessarily a real resource cost.

Linear contracts, defined by $c(x,a)$3, are computationally attractive but restrict the principal’s ability to tailor incentives across types and outcomes. Their worst-case multiplicative loss is linear in the number of types $c(x,a)$4, although they achieve a constant multiplicative approximation with an exponentially small additive loss. Designing near-optimal contracts is generally computationally hard: for every $c(x,a)$5, obtaining a multiplicative loss $c(x,a)$6 is NP-hard.

These results identify several possible TCA components:

• incentive and participation costs;
• agent rents;
• action inefficiency;
• adverse-selection loss from a common contract;
• contract-class loss from linearity;
• computational loss from tractability constraints.

They do not quantify negotiation, monitoring, enforcement, legal, administrative, or relationship costs.

Tax administration

A broad operational interpretation measures the resources required to transfer tax liabilities into public revenue. Let $c(x,a)$7 denote government administration, $c(x,a)$8 externally purchased taxpayer compliance, and $c(x,a)$9 internal taxpayer compliance. Total administrative cost is

$s\in\{0,1\}$0

with ratio

$s\in\{0,1\}$1

For Germany in 2021, the estimated tax revenue denominator was €832,091,000,000. Government tax-administration cost was estimated at €17,667,088,104, approximately 2.12% of revenue. Including taxpayer-side external and internal costs, the reported mean total administrative cost was €168,303,019,310, corresponding to 20.23% of revenue (Mantzaris et al., 2024).

The estimate is not a formal confidence interval. Internal taxpayer cost was imputed from an assumed 28% outsourcing rate:

$s\in\{0,1\}$2

The largest uncertainty therefore concerns the equivalence between external professional fees and the opportunity cost of internally performed compliance work. The paper identifies both likely undercounting, such as unmeasured internal tax departments and taxpayer time, and possible overcounting, such as non-tax services included in professional revenue.

The measure excludes spending-side administration, deadweight loss, legislative costs, and broader welfare distortions. It is consequently a measure of the operating and transaction cost of raising revenue, not a complete social cost of taxation.

Financial execution

Corporate-bond transaction cost analysis estimates the cost of trading in an over-the-counter market without a consolidated limit-order book. The cost is primarily decomposed into bid–ask-spread cost and price-impact or mid-price movement:

$s\in\{0,1\}$3

The study uses Enhanced TRACE data, identifies customer transaction signs, filters potential riskless-principal trades, infers transaction-based spreads from nearby opposite-sign trades, and estimates transient impact kernels. Customer buys have larger initial impact than customer sells, while buy impact decays slightly faster. The paper therefore supports side-specific execution benchmarks rather than a single symmetric spread benchmark (Guo et al., 2019).

A Bayesian execution-cost framework models implementation shortfall, VWAP, PWP20, and five-minute reversion using an asymmetric Laplace distribution with covariate-dependent location, scale, and skewness. The conditional expected benchmark is

$s\in\{0,1\}$4

The framework accounts for fat tails, skewness, heteroscedasticity, finite samples, and hierarchical partial pooling across broker algorithms (Markov, 2019).

These execution measures capture spread, timing, impact, adverse selection, participation, and reversion. They do not directly capture opportunity cost from unexecuted orders, explicit fees, operational costs, broker-selection costs, or the full cost of delegation. In a broader TCA account, execution benchmarks are components of agency cost rather than complete measures.

5. AI deployment, workflow, and human agency

Lifecycle deployment cost

The Levelized Cost of Artificial Intelligence (LCOAI) measures lifecycle AI cost per valid productive inference:

$s\in\{0,1\}$5

CAPEX includes infrastructure, training, fine-tuning, data, software, integration, security, compliance preparation, and initial labor. OPEX includes inference, API charges, cloud resources, storage, bandwidth, monitoring, maintenance, retraining, DevOps, security, and compliance.

The denominator excludes nonproductive calls such as health checks, background processes, and administrative queries. This distinction is important because raw inference volume does not equal useful agency. LCOAI is therefore a deployment-cost core for TCA, not a complete agency measure. It omits oversight, coordination, failure, governance, human displacement, and the value or reliability of actions (Curcio, 29 Aug 2025).

An illustrative one-year comparison at 10 million valid inferences reports LCOAI values of $s\in\{0,1\}$69.80 per 1,000 for Claude Haiku API, and $24.80 per 1,000 for a self-hosted LLaMA-2-13B deployment. The relative ordering changes with inference volume, CAPEX, OPEX, utilization, and infrastructure allocation. The study identifies an approximate self-hosting break-even region of 30–40 million annual inferences under its assumptions.

Multi-agent workflow attribution

For a DAG workflow $s\in\{0,1\}$7, TCA is defined as

$s\in\{0,1\}$8

The terms are:

• Base-prompt cost: tokens independently required by the node, including specialist prompts, tools, and user queries.
• Inference cost: output tokens generated by the node.
• Memory-injection cost: retrieved tokens inserted into the node’s prompt.
• Miss penalty: cost of falling back from a warm memory tier to a durable store.
• Context-accumulation cost: upstream context carried forward as workflow depth increases.

The exact memory-injection count is measured by two tokenizer passes:

$s\in\{0,1\}$9

This distinguishes tokens handed to a node from tokens generated by it. In the experimental harness, miss penalty and context accumulation were zero as separately billed terms by construction. Memory injection accounted for 13.6% of variable cost actionable by compile-time optimization and approximately 12%, more precisely 11.8%, of full billed cost. Its share increased from a structural zero at depth one to 27.6% at depth six. Injected tokens grew approximately linearly over measured depths two through six, with $\Delta E$0; a quadratic fit had a negative leading coefficient and did not demonstrate convex growth (Singh et al., 20 Sep 2026).

Reducing retrieval capacity from 32 to 2 entries lowered injected tokens by 28.7% and reduced measured task cost by 6.7% in the reported single-seed experiment, while accuracy changed from 0.600 to 0.570 within the stated seed-level variation. Graph rewrites were approximately cost-neutral in isolation, while model-tier assignment remained the dominant cost lever.

Coding-agent economics

An enterprise coding-agent study compares a cached frontier API configuration with a shared on-premise quantized model. It incorporates direct inference cost, infrastructure allocation, operations, developer repair labor, and quality effects. The API configuration incurred $\Delta E$12,560 and dedicated allocation cost $10,752. The local system incurred higher repair burden: Fix Commit Ratio was 74.93% compared with 45.93% for the API configuration, with Mantel–Haenszel repair odds ratio 3.61 after difficulty stratification.

Using a 15-minute repair assumption and a developer rate of $\Delta E$2988.75 for the API configuration and $\Delta E$39,773.96 for the API configuration, $\Delta E$414,051.25 for dedicated local deployment. Shared local deployment therefore saved 40.1% relative to the cached API configuration, while dedicated local deployment cost 43.8% more (Peng et al., 13 Jul 2026).

The study emphasizes that lower token or GPU cost does not necessarily imply lower TCA. A broader economic account includes:

$\Delta E$5

The empirical design is limited to one developer and two contiguous, non-randomized 28-day periods. Model, harness, quantization, serving stack, and period are confounded. Repair commits are not equivalent to escaped production defects, and developer-time proxies are not direct time-and-motion measurements. The strongest conclusion is conditional: cached frontier APIs can have low marginal inference cost, while shared on-premise infrastructure can have lower period-level TCO when fixed capacity is pooled; model quality, utilization, allocation, and rework determine the resulting cost frontier.

Human agency in AI-mediated communication

Human agency itself can have a measurable performance cost. In a study of 45 dyads involving non-native English speakers and native English speakers, participants used one of three machine-translation interfaces: quality labeling, regular post-editing, or augmented post-editing with LLM-generated paraphrase hints (Xiao et al., 11 Mar 2025).

Post-editing increased non-native speakers’ perceived control over message meaning. Mean agency ratings were 4.34 for labeling, 5.95 for regular post-editing, and 6.13 for augmented post-editing. However, augmented post-editing did not significantly increase agency over regular post-editing.

The agency benefit was accompanied by a communication trade-off. Post-editing produced more distinct information initiations but less elaboration and lower dyadic alignment. Rank-Biased Overlap alignment was 0.57 for labeling, 0.34 for regular post-editing, and 0.33 for augmented post-editing. The study did not find a significant workload difference, and it did not directly measure editing time, latency, keystrokes, verification effort, or cognitive load.

The empirical result is therefore not that agency is harmful in general. It demonstrates an agency–performance trade-off specific to the task and interface:

$\Delta E$6

A TCA interpretation includes inspection, editing, verification, language-production, delay, coordination, misalignment, and performance costs. The paper directly measures increased breadth, reduced depth, and reduced alignment; cognitive, temporal, emotional, and provider-side costs remain partly theoretical or unmeasured.

A related theoretical account describes AI assistance as a cumulative transfer of autonomous function. It distinguishes acute state cognitive bandwidth from slowly changing structural capacity. Repeated delegation may produce a surrender threshold beyond which human re-entry becomes difficult, with recovery requiring cognitive quiet, deliberate re-engagement, restored authority, and sometimes retraining. The paper does not define TCA or provide a numerical cost function, but it supplies a dynamic model of dependence, competence erosion, re-entry cost, and responsibility allocation (Margondai et al., 11 Jun 2026).

6. Measurement, aggregation, and limitations

A comprehensive TCA framework must specify its boundary, unit of analysis, time horizon, benefit measure, and treatment of transfers. A practical decomposition can distinguish the following layers:

1. Physical and informational cost: sensing, measurement, memory, computation, and entropy production.
2. Operational cost: recurring state-action costs, infrastructure, energy, maintenance, and administrative work.
3. Delegation and incentive cost: compensation, rents, monitoring, contracting, enforcement, and adverse selection.
4. Execution cost: spread, impact, timing, reversion, failed execution, and opportunity cost.
5. Human cost: attention, editing, verification, rework, training, competence loss, and responsibility burden.
6. Organizational cost: coordination, governance, reliability, compliance, security, incident response, and change management.
7. Outcome loss: errors, misalignment, delay, non-completion, defects, and unauthorized or low-value action.

A generic lifecycle formulation may be written as

$\Delta E$7

where the terms must be defined for the application. This is an editorial synthesis, not a formula established by any single cited paper.

Units and denominators

Different studies use incompatible denominators:

• work or entropy per cycle;
• expected cost per episode or average cost per period;
• basis points per transaction;
• dollars per valid inference;
• dollars per task;
• dollars per delegated action;
• percentage of tax revenue;
• developer hours or repair commits;
• quality-adjusted successful tasks.

A TCA result is not interpretable without its denominator. In particular, raw model calls, tokens, transactions, or approvals may not correspond to productive, authorized, or successful agency. LCOAI addresses this by using valid productive inferences; a broader agency metric may require completed authorized tasks weighted by quality and reliability.

Cost versus transfer

Payments to agents, brokers, contractors, or taxpayers may be transfers rather than social resource losses. A principal-level TCA may count compensation as cost, while a social-welfare TCA may separate transfer payments from effort, administration, monitoring, and deadweight loss. The principal–agent literature makes this distinction explicit: agent compensation reduces principal utility, but total surplus remains reward minus action cost.

Uncertainty and causal interpretation

Several sources emphasize model uncertainty:

• internal tax compliance costs depend on imputation assumptions;
• bond execution regressions are predictive rather than causal;
• broker-algorithm comparisons are observational and subject to order-selection effects;
• AI deployment comparisons depend on utilization, pricing, model quality, and allocation;
• human–AI agency studies often measure perceived control rather than actual semantic control;
• transit and fleet studies identify descriptive relationships rather than causal effects;
• thermodynamic inequalities provide bounds rather than universal equalities.

TCA should therefore report sensitivity analysis, uncertainty intervals where available, and explicit distinctions between measured, inferred, and extrapolated components.

Common misconceptions

Agency is not costless because a control rule is logically simple. In the information-engine model, reversible feedback can have zero direct entropy production while measurement remains energetically costly.

Mutual information is not itself a complete cost measure. Information bounds extracted work and lower-bounds measurement expenditure, but the model does not establish a universal equality.

Total cost is not necessarily total social cost. Administrative cost, transaction cost, lifecycle cost, and TCA each omit different categories.

Lower inference price is not lower agency cost. Infrastructure allocation, repair, oversight, quality, and opportunity costs can reverse the ranking of deployment alternatives.

More user control does not guarantee better outcomes. Post-editing increased perceived agency but reduced communication alignment in the reported translation study.

A human approval step does not guarantee meaningful human agency. The autonomy-surrender framework distinguishes formal authority from actual capacity to understand, evaluate, intervene, and assume responsibility.

A one-for-one replacement assumption can understate transition cost. In bus electrification, replacement ratios ranged from 1.101 for CTA to 1.245 for KAT, showing that operational feasibility can require more electric vehicles than displaced diesel vehicles (Bhidya et al., 31 Aug 2026).

Research directions

A mature TCA theory would require:

• a shared ontology for agency, benefit, cost, and responsibility;
• multi-level accounting from physical operation to organizational governance;
• quality-, reliability-, and authorization-adjusted denominators;
• explicit separation of transfers from real resource costs;
• dynamic models of dependence, competence, and recovery;
• causal designs for comparing delegated systems;
• uncertainty propagation across cost components;
• measurement of failed, delayed, and unexecuted actions;
• integration of safety, environmental, legal, and distributional effects;
• and domain-specific calibration of what constitutes successful agency.

The current literature supports a general conclusion: agency is an embodied, operational, and organizational accomplishment. Whether agency is realized by a one-bit thermodynamic controller, a stationary policy, a contract, a trading algorithm, a tax system, a transit fleet, or a multi-agent LLM workflow, its benefits are accompanied by costs of sensing, computation, coordination, actuation, oversight, correction, and responsibility. TCA is most useful when those costs are made explicit, assigned to the appropriate actors and time scales, and evaluated against the quality and legitimacy of the resulting action.