Every agentic system begins with an act of delegation. Before the first algorithm runs, before the first decision is made, before the first action is taken, a human being must make a choice: to entrust a portion of their authority, their judgment, and their agency to an autonomous system. This act of delegation is the foundational gesture of the agentic era. It is also, if poorly designed, the original sin from which all subsequent failures flow. How that delegation is designed - its scope, its boundaries, its conditions, and its mechanisms for revocation - determines whether the system that follows will serve its principal faithfully or subvert them incrementally.
The concept of delegation is not new. It is as old as human organisation itself. Every act of management, every appointment of an agent, every granting of power of attorney is an act of delegation. What is new is the nature of the delegate. When we delegate to a human being, we delegate to an entity that shares our embodied experience of the world, our social norms, our capacity for moral reasoning, and our vulnerability to consequence. When we delegate to an autonomous system, we delegate to an entity that possesses none of these things. The grammar of delegation - the rules, structures, and conventions that govern how authority is transferred, exercised, and reclaimed - must therefore be fundamentally redesigned for the agentic context.
This essay explores the principles and practices of Delegation Design - the discipline of creating the structures through which human authority is transferred to, exercised by, and recovered from autonomous agents. It is a discipline that sits at the intersection of interaction design, contract law, organisational theory, and moral philosophy. It is, we shall argue, the most consequential design discipline of the agentic era, for it is the discipline that determines the terms of the compact between human and machine.
01 The Principal-Agent Compact
The relationship between a human and an autonomous agent is, at its core, a principal-agent relationship. The principal (the human) delegates authority to the agent (the autonomous system) to act on their behalf. This relationship is governed by a compact - a set of mutual expectations, obligations, and constraints that define the terms of the delegation. In traditional principal-agent relationships, this compact is shaped by legal frameworks, social norms, professional codes, and the shared understanding that comes from being members of the same species. In the agentic context, most of these scaffoldings are absent. The compact must be designed from first principles - the agentic design principles that govern every human-agent relationship.
The principal-agent compact in agentic systems has four essential components. First, scope - what the agent is authorised to do and, equally importantly, what it is not authorised to do. Second, constraints - the rules, limits, and conditions under which the agent must operate. Third, accountability - the mechanisms by which the agent reports its actions and the principal verifies them. Fourth, revocability - the means by which the principal can modify, suspend, or terminate the delegation. Each of these components must be explicitly designed, clearly communicated, and reliably enforced.
The challenge is that this compact must be both precise enough to prevent misuse and flexible enough to allow the agent to exercise the judgment that makes it useful. A delegation that is too tightly constrained reduces the agent to a simple automation - a script with extra steps. A delegation that is too loosely constrained creates a rogue agent - a system that acts with authority but without adequate guidance. The art of Delegation Design is the art of finding the productive middle ground between these extremes.
Delegation without design is abdication. The act of giving authority to a machine is not a surrender of responsibility - it is an assumption of a higher form of it.
02 The Grammar of Authority
Authority, in the context of Delegation Design, is not a binary state. It is not simply "on" or "off," "granted" or "withheld." It is a structured, multi-dimensional space with its own grammar - its own syntax and semantics for expressing what an agent can do, under what conditions, and with what consequences. Designing this grammar is one of the most important tasks in Agentic Experience Design.
The grammar of authority has several key dimensions. The first is action scope - the set of actions the agent is permitted to take. This is not simply a list of allowed operations; it is a structured hierarchy of capabilities, from the most routine (checking a price) to the most consequential (executing a purchase). Each level of the hierarchy carries different implications for risk, reversibility, and the need for human confirmation.
The second dimension is resource scope - the resources the agent is permitted to use. This includes financial resources (spending limits), informational resources (what data the agent can access), and relational resources (who the agent can contact or negotiate with on the principal's behalf). Resource scope is particularly critical in Agentic Commerce, where the agent may be authorised to spend real money.
The third dimension is temporal scope - the time period during which the delegation is active. Some delegations are permanent ("always manage my email inbox"). Some are time-bounded ("book my travel for next week's conference"). Some are event-triggered ("if the stock drops below $50, sell"). The temporal dimension interacts with the concept of Temporal Trust - the way trust evolves over time through accumulated experience.
The fourth dimension is contextual scope - the conditions under which the delegation applies. An agent might be authorised to book flights under $500 for domestic travel, but require approval for international flights regardless of price. The contextual dimension requires the agent to assess not just what it can do, but whether the current situation falls within the conditions under which it should do it. This is the domain of the Operational Envelope.
03 Delegation Boundaries
The boundaries of a delegation are not walls; they are membranes. They must be permeable enough to allow the agent to operate effectively, but selective enough to prevent it from exceeding its authority. The design of these boundaries is one of the most delicate tasks in Delegation Design, requiring a deep understanding of both the agent's capabilities and the principal's risk tolerance.
There are two fundamental approaches to boundary design. The first is prescriptive boundaries - explicit rules that define what the agent can and cannot do. "Do not spend more than $200 on any single purchase." "Do not schedule meetings before 9am." "Do not sell any stock that has been held for less than 30 days." Prescriptive boundaries are clear, enforceable, and easy to audit. They are also brittle - they cannot anticipate every situation, and they can prevent the agent from taking beneficial actions that fall outside their scope.
The second approach is principled boundaries - higher-level guidelines that express the principal's values and priorities, leaving the agent to interpret them in context. "Prioritise value over price." "Respect my work-life balance." "Act conservatively with my investments." Principled boundaries are more flexible and more resilient to novel situations. They are also more ambiguous, harder to enforce, and more susceptible to misinterpretation. They require the agent to possess a sophisticated model of the principal's intent - what we call Intent Architecture.
In practice, effective Delegation Design uses a combination of both approaches. Prescriptive boundaries define the hard limits - the lines that must never be crossed. Principled boundaries guide the agent's judgment within those limits. The interaction between the two creates a structured space of authorised action - a space that is bounded but not rigid, constrained but not paralysed. This structured space is the Delegation Scope.
The best delegation boundaries are like the banks of a river - they do not stop the water from flowing; they give it direction, purpose, and power.
04 Progressive Delegation
Trust is not granted; it is earned. And the earning of trust is a gradual process, built through repeated demonstrations of competence, reliability, and integrity. Progressive delegation is the design pattern that mirrors this process - it begins with a narrow scope of authority and expands it incrementally as the agent demonstrates its trustworthiness.
The progressive delegation model has several stages. In the observation stage, the agent watches the principal perform tasks and learns their preferences, patterns, and priorities. It makes no autonomous decisions; it only observes and models. In the suggestion stage, the agent begins to offer recommendations - "Based on your usual preferences, I would suggest booking the 8am flight." The principal accepts or rejects these suggestions, and the agent refines its model accordingly.
In the confirmation stage, the agent begins to take action, but only with explicit approval. "I have found a flight that matches your preferences. Shall I book it?" The principal reviews and approves each action, building confidence in the agent's judgment. In the notification stage, the agent acts autonomously within defined boundaries and notifies the principal after the fact. "I have booked your flight for tomorrow. Here are the details." The principal retains the ability to override or reverse the action, but the default has shifted from "ask first" to "act first."
Finally, in the autonomous stage, the agent operates with full authority within its Delegation Scope, reporting only exceptions and significant decisions. The principal trusts the agent to handle routine matters without intervention, freeing their attention for higher-order concerns. This is the stage at which the agent becomes truly useful - not as a tool that must be operated, but as a partner that can be relied upon.
The progression through these stages is not automatic; it is designed. It is governed by the Interrupt Frequency - the rate at which the agent seeks human input. As trust builds, the interrupt frequency decreases. As the agent's scope expands, the interrupt frequency may temporarily increase to allow the principal to calibrate their comfort with the new level of authority. The design of this progression - its pace, its triggers, its feedback mechanisms - is one of the most important aspects of Delegation Design.
05 The Revocability Imperative
No delegation should be irrevocable. This is not merely a design principle; it is a moral imperative. The ability to reclaim delegated authority - to narrow the scope, to increase the constraints, to suspend the delegation entirely - is the ultimate safeguard of human agency in the agentic era. Effective agent oversight mechanisms make this revocability practical rather than theoretical. A system that makes delegation easy but revocation difficult is a system that traps its users in relationships they cannot control.
Revocability operates at multiple levels. Immediate revocation is the ability to stop the agent from taking a specific action in real-time - a "cancel" button, a voice command, a gesture. This is the most basic form of revocation, and it must be available at all times, regardless of the agent's current state or the complexity of the task it is performing. Scope revocation is the ability to narrow the agent's authority - to remove certain actions from its repertoire, to lower its spending limits, to restrict its access to certain resources. This is a more nuanced form of revocation, and it requires a well-designed interface for managing the delegation's parameters.
Complete revocation is the ability to terminate the delegation entirely - to return all authority to the principal and to ensure that the agent takes no further autonomous action. This is the nuclear option, and it should be available but rarely needed. Its very existence, however, is essential to the principal's sense of control and, therefore, to their willingness to delegate in the first place. The knowledge that one can revoke is what makes one willing to delegate.
The design of revocation mechanisms must account for the consequences of revocation. When a delegation is revoked mid-task, the agent may be in the middle of a complex, multi-step operation. What happens to the steps that have already been completed? What happens to the commitments that have already been made? A well-designed revocation mechanism includes a wind-down protocol - a structured process for gracefully terminating the agent's activities, honouring existing commitments where possible, and clearly communicating the state of affairs to the principal. This is where Delegation Design intersects with Failure Architecture - the design of graceful endings.
The power to delegate is meaningful only when it is accompanied by the power to revoke. Authority without revocability is not delegation - it is surrender.
06 Delegation in Agentic Commerce
The principles of Delegation Design take on heightened significance in the context of Agentic Commerce - the emerging domain in which autonomous agents act as economic participants, buying, selling, negotiating, and transacting on behalf of human principals. In agentic commerce, the stakes of delegation are measured in currency, and the consequences of poorly designed delegation are financial.
A shopping agent, for example, must operate within a carefully designed delegation that specifies not only what it can buy and how much it can spend, but also the criteria by which it evaluates products, the merchants it is permitted to transact with, the payment methods it can use, and the conditions under which it should seek human approval. The agent checkout - the moment when an agent commits to a transaction - is the point where delegation design meets commercial reality. Each of these parameters represents a dimension of the delegation's grammar, and each must be calibrated to the principal's preferences and risk tolerance.
The concept of spending authority is particularly critical. A naive approach might simply set a spending limit - "do not spend more than $500 per month." But a well-designed delegation would express spending authority in more nuanced terms: per-transaction limits, per-category limits, approval thresholds that vary by product type, and dynamic limits that adjust based on the agent's track record. The economics of trust in agentic systems means that these spending constraints carry real financial weight - poorly calibrated delegation boundaries create measurable economic loss. The MIT Media Lab's research on authenticated, authorised, and auditable delegation of authority to AI agents provides a useful framework for thinking about these structures.
Agentic commerce also introduces the challenge of multi-party delegation. When an agent negotiates with another agent on behalf of its principal, both agents are operating under delegations from their respective principals. The negotiation is not just a transaction; it is an interaction between two delegation frameworks. The design of these interactions - the protocols for establishing mutual trust, verifying authority, and resolving disputes - is one of the frontier challenges of the field. The question of liability when agents fail becomes acute in multi-party delegation, where the chain of authority spans multiple principals and multiple agents.
07 The Moral Weight of Delegation
Delegation is not a morally neutral act. When we delegate authority to an autonomous system, we are making a moral choice - a choice about what we are willing to entrust to a machine, what values we expect it to uphold, and what consequences we are prepared to accept. The design of delegation is therefore a moral design, and it carries moral weight.
The Ada Lovelace Institute's work on "The Dilemmas of Delegation" identifies several moral tensions inherent in the act of delegating to AI agents. The first is the tension between efficiency and accountability. Delegation to autonomous systems can dramatically increase efficiency, but it can also diffuse accountability. When an agent makes a harmful decision, who is responsible - the agent, the principal who delegated to it, or the designer who built it? A well-designed delegation framework must address this question explicitly, establishing clear lines of accountability that survive the act of delegation.
The second tension is between autonomy and control. The value of an autonomous agent lies precisely in its ability to act independently - to make decisions, to exercise judgment, to adapt to circumstances. But this autonomy comes at the cost of control. The more autonomous the agent, the less the principal can predict or direct its behaviour. Delegation Design must navigate this tension, finding the level of autonomy that maximises value while maintaining an acceptable level of control.
The third tension is between convenience and consent. It is tempting to design delegation systems that are frictionless - that make it easy and effortless to grant authority to an agent. But frictionless delegation risks undermining informed consent. If the principal does not fully understand what they are delegating, the delegation is not truly voluntary. This is the domain of the Consent Horizon - the boundary between what the user has knowingly authorised and what the system assumes on their behalf. Delegation Design must ensure that this boundary is clear, visible, and respected.
08 Conclusion: The Architecture of Trust Transfer
Delegation is the foundational act of the agentic era. It is the moment at which human authority flows into the machine, the moment at which trust is transferred from the principal to the agent. The quality of this moment - its clarity, its intentionality, its respect for the principal's autonomy - determines the quality of everything that follows.
Delegation Design is the discipline that shapes this moment. It provides the grammar for expressing authority, the structures for bounding it, the patterns for expanding it progressively, and the mechanisms for reclaiming it when necessary. It is a discipline that draws on interaction design, contract law, organisational theory, and moral philosophy. It is, in its essence, the design of the relationship between human and machine - not the interface through which they interact, but the compact that governs their collaboration.
The stakes of this discipline could not be higher. As autonomous agents become more capable and more prevalent, the acts of delegation that govern them will shape the distribution of power, the allocation of responsibility, and the nature of human agency itself. If we design delegation well, we create a future in which human authority is amplified by machine capability - a future of genuine partnership. If we design it poorly, we create a future in which human authority is eroded by machine autonomy - a future of gradual, comfortable surrender. The choice is ours, and it begins with the design of the first delegation.
Tony Wood
Founder of the AXD Institute and a leading voice in the field of Agentic Experience Design.