The most dangerous AI agent is not the one that makes bad decisions. It is the one whose decisions cannot be read. An agent that makes a poor choice within a legible framework can be corrected, constrained, or redirected. An agent whose reasoning is opaque - whose identity is unclear, whose decision process is invisible, whose outcomes are reported without context - cannot be trusted, cannot be calibrated, and cannot be governed. This is the legibility crisis of the agentic age, and it is the subject of this essay.
Agent legibility is the quality of an autonomous AI system that makes its identity, capabilities, constraints, decision-making processes, and operational state structurally readable - both to the humans who delegate to it and to the other agents and systems it interacts with. It is the bridge between Signal Clarity (how merchants make themselves readable to agents) and trust architecture (how trust is formed, calibrated, and maintained in human-agent relationships). Without agent legibility, neither Signal Clarity nor trust architecture can function.
The term “agent legibility” is emerging in practitioner discourse as the operational expression of what the AXD Institute has called Signal Clarity from the agent’s perspective. This essay examines what agent legibility actually means, why it matters for agentic commerce, and how it connects to the broader architecture of Agentic Experience Design.
01 - The Legibility Crisis
The current generation of AI agents operates in a condition of profound illegibility. When a customer delegates a task to an AI shopping agent, they typically have no structured way to understand what the agent is, who built it, what data it accesses, what constraints govern its behaviour, how it makes decisions, or how it will report its outcomes. The agent is a black box that accepts instructions and returns results. The space between instruction and result - the space where all the consequential decisions happen - is invisible.
This illegibility is not a technical limitation. It is a design failure. The technology to make agents legible already exists: structured identity declarations, decision audit trails, outcome reporting frameworks, capability manifests. What does not yet exist is the design discipline that specifies how these technical capabilities should be composed into a coherent legibility architecture. This is the gap that Agentic Experience Design exists to fill.
The legibility crisis is not merely an inconvenience. It is a structural barrier to the adoption of agentic commerce. Research consistently shows that the primary obstacle to AI agent adoption is not capability but trust. Customers are willing to delegate to agents that they can understand and evaluate. They are unwilling to delegate to agents that operate as opaque systems. Agent legibility is not a nice-to-have feature - it is the precondition for commercial viability.
02 - What Agent Legibility Actually Means
Agent legibility is not transparency in the traditional sense. Transparency, as commonly understood in AI ethics discourse, means “showing the user what is happening.” It typically manifests as explanability features: “Here is why the algorithm made this recommendation.” This is necessary but insufficient for agentic systems, because agentic systems do not merely recommend - they act. The user needs to understand not just what the agent thinks but what it does, what it is authorised to do, and what it will do next.
Agent legibility, as the AXD Institute defines it, is structural readability - the property of a system whose behaviour can be understood, predicted, and evaluated by external observers without requiring access to its internal state. A legible agent is one whose identity is declared, whose capabilities are specified, whose constraints are visible, whose decision process is auditable, and whose outcomes are reported in terms that the delegating human can evaluate against their original intent.
This definition has two audiences. First, human legibility: the customer who delegated to the agent must be able to read its behaviour in terms of their own goals, values, and expectations. Second, agent-to-agent legibility: other agents in the ecosystem must be able to read the agent’s identity, authority, and constraints in order to interact with it appropriately. A shopping agent that interacts with a merchant’s agent must be able to verify the merchant agent’s identity, understand its authority, and assess its reliability. This bidirectional legibility is the communication infrastructure of agentic commerce.
03 - Signal Clarity and the Legibility Stack
The relationship between agent legibility and Signal Clarity is complementary. Signal Clarity, the first pillar of the Five Pillars of AXD Readiness, addresses how merchants make their products and services readable to agents. It is the merchant-side legibility obligation: structured data, machine-readable product attributes, verifiable quality claims, and agent-consumable content. Signal Clarity answers the question: “Can the agent understand what the merchant is offering?”
Agent legibility addresses the complementary question: “Can the human understand what the agent is doing?” Together, Signal Clarity and agent legibility form what the AXD Institute calls the legibility stack - the complete communication infrastructure that enables trust-governed commerce between humans, agents, and merchants. The legibility stack has three layers: merchant-to-agent legibility (Signal Clarity), agent-to-human legibility (outcome reporting and trust calibration), and agent-to-agent legibility (identity verification and capability exchange).
The practitioner community has begun using the term “agent legibility” to describe what the AXD Institute originally framed as the agent-facing dimension of Signal Clarity. This terminological evolution is welcome: it gives practitioners a concrete, actionable concept that maps directly to the design challenges they face. The AXD Institute adopts the term as a bridging concept that connects the theoretical framework of Signal Clarity to the practical requirements of agent design.
04 - The Three Layers of Agent Legibility
Agent legibility operates at three distinct layers, each addressing a different dimension of the readability challenge.
Identity legibility is the foundation. The agent declares what it is, who it represents, what authority it holds, and what constraints govern its behaviour. In agentic commerce, identity legibility means the shopping agent can declare: “I am Agent X, acting on behalf of Customer Y, with delegated authority to purchase household supplies within a monthly budget of £200, constrained by the customer’s dietary preferences and brand exclusions.” This declaration enables merchants, other agents, and regulatory systems to interact with the agent appropriately.
Process legibility is the middle layer. The agent makes its decision-making process structurally readable, so that observers can understand not just what it decided but how and why. Process legibility does not require exposing the agent’s internal model weights or algorithmic details. It requires structured accounts of the factors considered, the alternatives evaluated, the trade-offs made, and the criteria applied. The customer who delegated “find me a good hotel in Paris” should be able to see that the agent considered 47 hotels, filtered by the customer’s stated preferences, ranked by a weighted combination of price, location, and reviews, and selected the option that best matched the outcome specification.
Outcome legibility is the surface layer. The agent provides clear, comprehensible accounts of what it did, what resulted, and how the outcome relates to the original delegation. Outcome legibility is what the customer actually sees - the report that arrives after the agent has acted. A legible outcome report does not merely state what was purchased. It explains why this option was selected, what alternatives were considered, what trade-offs were made, and how the outcome serves the customer’s stated goals. This is the information the customer needs to calibrate their trust in the agent for future delegations.
05 - Legibility as Trust Infrastructure
Agent legibility is not a feature. It is infrastructure. Specifically, it is the infrastructure that enables trust architecture to function. Without legibility, trust cannot be formed, calibrated, or recovered - because the human has no basis for evaluating the agent’s behaviour.
Trust formation requires identity legibility: the customer must be able to understand what the agent is and what it is authorised to do before they delegate. Trust calibration requires process and outcome legibility: the customer must be able to evaluate the agent’s decisions and outcomes in order to adjust their delegation - expanding scope when the agent performs well, constraining it when it does not. Trust recovery requires all three layers: when something goes wrong, the customer must be able to understand what happened (outcome legibility), why it happened (process legibility), and whether the failure was within the agent’s declared capabilities (identity legibility).
This is why the AXD Institute treats agent legibility as a structural requirement rather than a feature request. An agent without legibility is an agent that cannot participate in trust-governed commerce. It may be technically capable of executing transactions, but it cannot sustain the human-agent relationship that makes ongoing delegation viable. Legibility is the bridge between capability and trust.
06 - The Merchant’s Legibility Obligation
Agent legibility is not solely the responsibility of agent developers. Merchants have a complementary legibility obligation: they must make their products, services, and commercial terms legible to the agents that will evaluate and transact with them. This is the merchant-facing dimension of the legibility stack - what the AXD Institute has previously described as Signal Clarity.
The merchant’s legibility obligation includes structured product data (machine-readable attributes, not just human-readable descriptions), verifiable quality claims (backed by data, not just marketing assertions), transparent pricing (including all conditions, constraints, and dynamic pricing rules), and compliance declarations (regulatory status, ethical sourcing, environmental impact). Each of these elements must be structured for agent consumption - not as a supplement to the human-facing product page, but as a parallel communication channel designed specifically for machine customers.
The merchant who fulfils their legibility obligation gains a structural advantage in agentic commerce. Their products are more likely to be discovered, evaluated, and selected by agents, because agents can process and compare their offerings more effectively. The merchant who does not - whose products are described only in human-readable marketing copy, whose quality claims are unverifiable, whose pricing is opaque - will be systematically disadvantaged as the proportion of agent-mediated commerce grows.
07 - Designing for Machine Readers
The design of agent legibility requires a fundamental shift in how we think about communication in commercial systems. Traditional design communicates with humans through visual interfaces - layouts, typography, colour, imagery. Agent legibility communicates with machines through structured data - schemas, protocols, APIs, and verifiable declarations.
This does not mean that agent legibility is purely a technical challenge. The design decisions involved are deeply consequential: What information should an agent declare about itself? How much process detail is useful versus overwhelming? What level of outcome reporting enables trust calibration without creating information fatigue? How should legibility requirements differ for high-stakes versus low-stakes delegations? These are design questions, not engineering questions, and they require the kind of disciplined thinking that Agentic Experience Design provides.
The AXD Institute proposes that agent legibility should be designed as a graduated system - one that provides different levels of detail for different audiences and contexts. The customer who delegates a routine coffee purchase needs minimal process legibility: “I bought your usual brand at the usual price.” The customer who delegates a complex travel booking needs extensive process legibility: “I evaluated 47 options across 12 criteria, here are the top three alternatives I considered, and here is why I selected this one.” The legibility architecture must adapt to the complexity and stakes of the delegation.
08 - The Legibility Imperative
Agent legibility is not optional. It is the structural precondition for every other element of Agentic Experience Design. Without legibility, delegation design cannot function - because the customer cannot evaluate whether the agent honoured their delegation. Without legibility, trust architecture cannot function - because trust cannot be calibrated without observable behaviour. Without legibility, the Five Pillars of AXD Readiness cannot be assessed - because readiness requires measurable, verifiable properties that depend on structural readability.
The organisations that will lead in agentic commerce are those that treat legibility as a first-class design requirement - not an afterthought, not a compliance checkbox, but a structural property of every agent they build and every commercial system they operate. The legibility imperative is not about making agents transparent for the sake of transparency. It is about making agents readable for the sake of trust - and trust is the primary material of the agentic age.
The AXD Institute will continue to develop the conceptual architecture of agent legibility - refining the three-layer model, specifying design patterns for graduated legibility, and building practical frameworks that organisations can use to assess and improve the legibility of their agentic systems. Agent legibility is not a destination. It is a design practice - one that must evolve as the capabilities and complexity of autonomous agents continue to grow.