top of page

Agentic Governance: Is It Time to Replace Principal-Agent Theory with Agent-Agent Theory?

  • Writer: Michael Hilb
    Michael Hilb
  • May 31
  • 7 min read

Artificial intelligence is beginning to reshape not only how organizations make decisions, but also how governance itself is exercised. Much of the current discussion focuses on the extent to which artificial intelligence can support, augment, or automate human decision-making. Yet the emergence of agentic intelligence raises a more fundamental question: What happens when artificial systems increasingly interact, negotiate, decide, and act autonomously with other artificial systems?



The Emergence of Agentic Governance


Agentic governance adds a new dimension to artificial governance. Artificial governance broadly refers to the application of artificial intelligence to governance processes. Artificial systems may collect information, identify patterns, generate recommendations, evaluate alternatives, or even execute predefined decisions.


Agentic governance represents a qualitatively different configuration. It emerges when artificial agents are delegated the capacity not only to process information, but also to pursue objectives, interact with other agents, negotiate trade-offs, make decisions, and initiate actions without continuous human intervention. The artificial system thereby moves from being predominantly an instrument of governance toward becoming an actor within the governance relationship itself.


This changes the central question. Artificial governance asks which governance decisions can and should be automated. Agentic governance asks which governance interactions can and should be delegated to autonomous agents.


The Boundaries of Agentic Governance


The future scope of agentic governance will ultimately depend on several interrelated boundaries: technological feasibility, economic desirability, legal accountability, and ethical responsibility.


Technological Feasibility

From a technological perspective, the limits of agentic governance are likely to become progressively less significant. In increasingly connected organizational environments, artificial agents can theoretically be given access to information, resources, transactional systems, decision rights, and communication channels.


As artificial intelligence becomes more capable and organizational systems become more interconnected, the technological question may therefore increasingly shift from whether agentic governance is possible to where organizations and societies choose to impose boundaries upon it.


Economic Desirability

The economic logic is similarly compelling. Artificial agents may process information continuously, coordinate at machine speed, reduce transaction costs, and apply decision rules more consistently than human actors. Organizations will consequently have strong incentives to delegate governance interactions wherever automation improves efficiency, responsiveness, or cost effectiveness.


Yet governance cannot be reduced to efficiency. Governance is also concerned with the allocation of power, the reconciliation of competing interests, the exercise of judgment, the establishment of legitimacy, and the assignment of accountability. A governance interaction may therefore be highly efficient without necessarily being desirable.


The relevant principle is thus unlikely to be maximal automation. A more plausible model is bounded delegation: governance interactions may increasingly be delegated to artificial agents where the benefits are substantial, but only within boundaries that preserve appropriate levels of legal accountability and ethical responsibility.


Legal Accountability


The legal dimension may ultimately prove even more consequential. Corporate law is built on the concept of the corporation as a legal person whose actions are ultimately attributable to human directors and officers. Agentic governance challenges this architecture because it increasingly separates operational decision-making from direct human agency.


The central legal question is therefore not simply whether an artificial agent may execute an action. Automated systems already do so extensively. The more fundamental question is whether an artificial agent could become an independent bearer of authority, responsibility, liability, or even legal personality.


Two broad schools of thought have emerged. A more technocratic perspective favors attributing responsibility directly to artificial agents, thereby reducing or limiting the liability of those who designed, appointed, deployed, or supervised them.


A more accountability-oriented perspective moves in the opposite direction, seeking to broaden rather than dilute responsibility by creating legal constructs through which obligations, assets, liabilities, and responsibilities could be attributed to artificial agents without removing the accountability of the human actors behind them.


Which approach ultimately prevails will be shaped as much by politics and political philosophy as by technology itself. Different political systems, grounded in different conceptions of authority, legitimacy, and individual autonomy, are likely to assess the trade-offs very differently.


Yet across political systems, it seems unlikely that societies will fully surrender ultimate authority to artificial agents. Authoritarian regimes may embrace such agents as instruments of efficiency, surveillance, and control, while resisting institutional arrangements that dilute sovereign human authority.


Democratic societies may arrive at a similar boundary for very different reasons: citizens are unlikely to accept the wholesale removal of human agency where decisions materially affect employment, wealth, rights, or other fundamental social interests.


The underlying rationale may differ, but the outcome could be similar: artificial agents may exercise extensive delegated authority without becoming the ultimate source of legitimate authority.


The more plausible development is therefore a layered allocation of agency. Artificial agents may increasingly exercise substantial authority at lower organizational levels, while human beings remain the ultimate bearers of constitutional, fiduciary, and political accountability.


Ethical Responsibility


The legal question itself rests on a deeper philosophical issue: What does it mean to possess agency? Agency is generally associated with the ability to form intentions, make choices, pursue objectives, and bear responsibility for resulting actions. This appears intuitively to distinguish humans from machines. Yet the distinction becomes less straightforward under closer examination.


Artificial agents operate on the basis of architecture, training data, optimization functions, environmental inputs, and institutional constraints. Human beings, in turn, act on the basis of biological predispositions, socialization, experience, incentives, values, institutional environments, and neurological processes.

The comparison therefore leads directly to one of philosophy's oldest unresolved questions: the problem of free will.


An artificial system may increasingly behave as though it possesses agency without society accepting that it possesses moral agency in the same sense as a human being. The distinction between functional agency and moral agency may consequently become one of the defining questions of future governance systems.


From Principal-Agent to Agent-Agent Theory?


Even if societies continue to insist that humans remain the ultimate bearers of accountability, agentic governance is likely to emerge extensively at lower levels of organizations.


This raises an important theoretical question. If artificial agents increasingly interact directly with other artificial agents, does one of the foundational concepts of corporate governance – principal-agent theory – need to be replaced by an agent-agent theory?


At first sight, the proposition appears plausible. Principal-agent theory traditionally examines situations in which a principal delegates authority to an agent whose actions cannot be perfectly observed and whose interests may diverge from those of the principal. Governance mechanisms such as monitoring, incentives, reporting systems, contractual safeguards, and boards are designed to mitigate the resulting agency problems.


With agentic governance, the visible structure of these relationships begins to change. Governance moves from predominantly human-human relationships toward combinations of human-agent and agent-agent relationships.


Yet this does not necessarily invalidate principal-agent theory. The principal-agent relationship is not fundamentally defined by the biological identity of the actors involved. Its core lies in delegation under conditions of imperfect information and potentially divergent objectives. Those conditions do not disappear when the agent is artificial. They may, in fact, become even more important.


From Interest Alignment to Objective Alignment


Traditional principal-agent theory frequently assumes that agents possess some form of self-interest. Human managers may seek higher compensation, status, security, influence, or private benefits.


Artificial agents do not possess self-interest in this psychological sense. They do not seek larger offices, prestigious titles, or higher social status. But this does not mean that artificial agents are free from the functional equivalent of conflicting interests.


The functional equivalent of human self-interest is not machine desire, but objective-function specificity. Agentic governance does not eliminate the problem of interest divergence. It transforms it into a problem of objective divergence. The governance challenge consequently shifts from aligning interests to aligning objectives.


Hyper-Efficient Misalignment as a New Governance Risk


The issue is therefore no longer merely whether an agent behaves opportunistically, as traditionally emphasized in principal-agent theory. The deeper question becomes whether the objective being optimized is the right objective in the first place.


Artificial agents may eventually become much more effective at pursuing narrowly specified objectives with extraordinary consistency and persistence. This creates a distinctive governance risk: hyper-efficient misalignment – an agent pursuing the wrong objective exceptionally well.


A subsidiary may maximize its own profitability at the expense of the wider group. A procurement function may maximize cost efficiency at the expense of resilience. A financial agent may maximize returns at the expense of environmental or social outcomes. More generally, optimization at one level of a system does not necessarily produce optimization at the level of the system as a whole.


The fundamental tension is therefore between individual value optimization and collective value optimization. This becomes particularly important in multi-agent systems. If every artificial agent is designed to maximize the objective assigned to it, increasingly capable agents may reproduce familiar organizational conflicts with greater speed, consistency, and scale.


This insight reinforces rather than undermines the relevance of principal-agent theory. Its deepest proposition is not that human agents are selfish. It is that delegation creates the possibility that the delegated actor will optimize an objective that differs from that of the delegating actor – or from that of the wider system.


Toward a Layered Architecture of Human and Artificial Agency


Agentic governance adds an important new dimension to the evolution of corporate governance, but it does not require principal-agent theory to be replaced by agent-agent theory. Indeed, the rise of artificial agents may reveal that the principal-agent problem is more fundamental than its traditional application to human managers suggests. As artificial agents become increasingly capable of optimizing the objectives assigned to them, the central governance question shifts from who performs the optimization to who defines its purpose and boundaries.


The future of governance is therefore likely to take the form of a layered architecture in which human and artificial agency coexist and operate at different levels. Artificial agents may become extraordinarily effective at optimizing within a given framework. Governance at its highest level, however, remains a matter of constitutional choice: determining what should be optimized, whose interests matter, which trade-offs are legitimate, how far authority may be delegated, and where ultimate accountability resides.


The agentic age may therefore not mark the end of the principal-agent problem. On the contrary, it may elevate the role of the principal to unprecedented importance.


The author employed AI-based writing tools to support the drafting process. All core ideas, arguments, and conceptual contributions are solely those of the author.

Comments


bottom of page