Agent Legibility

Agent legibility is the quality of being readable. In traditional software, legibility meant clean code and clear documentation. In agentic AI, legibility means something far more consequential: it means that an autonomous agent's behaviour can be understood by the humans it serves, the institutions it interacts with, and the other agents it collaborates with. Agent legibility is not the same as AI explainability - explainability focuses on why a model made a specific prediction, while legibility encompasses the entire agent: its identity, its authority chain, its operational history, its current constraints, and its failure modes. An agent can be explainable (its model decisions are interpretable) without being legible (its overall behaviour, authority, and limitations are not readable). Legibility is the broader design requirement.

Agent legibility operates across three dimensions that must be designed together. Identity legibility answers the question 'who is this agent?' - it encompasses the agent's credentials, its authorising principal, its delegation chain, and its organisational affiliation. This is the domain of Know Your Agent (KYA) and Agent Registries. Action legibility answers the question 'what has this agent done and why?' - it encompasses audit trails, decision logs, reasoning traces, and post-hoc explanations. This is the domain of the AXD Explainability and Observability Design Standard. Capability legibility answers the question 'what can this agent do and what can it not do?' - it encompasses operational envelope documentation, constraint specifications, and limitation disclosures. This is the domain of the Onboarding and Capability Discovery Framework. All three dimensions must be present for an agent to be considered legible.

Agent legibility is not only a human-facing requirement. In multi-agent systems - where autonomous agents collaborate, negotiate, and transact with each other - machine legibility becomes critical. An agent negotiating a purchase must be able to read the selling agent's credentials, verify its authority chain, and evaluate its track record. This requires structured, machine-readable formats for agent identity (verifiable digital credentials), agent capabilities (structured capability manifests), and agent history (machine-readable performance records). Machine legibility is the foundation of agent-to-agent trust. Without it, every multi-agent interaction requires human intermediation, which defeats the purpose of autonomous operation. The IAB Tech Lab's Agent Registry and Google's Universal Commerce Protocol (UCP) are early infrastructure attempts to solve machine legibility at scale.

Agent legibility is the prerequisite for trust architecture. Trust cannot be formed in an agent whose actions are opaque, whose authority is unverifiable, and whose limitations are undisclosed. The AXD trust architecture model identifies three trust requirements that depend directly on legibility: competence trust (the human believes the agent can do what it claims - requires capability legibility), integrity trust (the human believes the agent will act within its authorised scope - requires action legibility), and benevolence trust (the human believes the agent acts in their interest - requires identity legibility showing the delegation chain). Trust calibration - the ongoing process of aligning human confidence with agent reliability - is impossible without legibility. If the human cannot read the agent's actions, they cannot calibrate their trust. If they cannot calibrate their trust, they will either over-trust (dangerous) or under-trust (wasteful).

Designing for agent legibility requires decisions at every layer of the agentic system. At the identity layer: implement verifiable digital credentials, publish delegation chains, and register with Agent Registries. At the action layer: maintain comprehensive audit trails, log all decisions with reasoning traces, and provide post-hoc explanations in both human-readable and machine-readable formats. At the capability layer: document the operational envelope (what the agent is designed to handle), publish constraint specifications (what the agent is not permitted to do), and disclose known limitations (where the agent is likely to fail). The AXD Practice provides three frameworks that directly address legibility: the Explainability and Observability Design Standard (action legibility), the Onboarding and Capability Discovery Framework (capability legibility), and the Multi-Agent Orchestration Visibility Model (machine legibility in multi-agent systems). Legibility is not a compliance checkbox - it is a design discipline that determines whether an agentic system can earn and maintain trust.