An agent capability is a describable functional competence or operational capacity of an agent, or in the threat-model reading a functional ability that an agent claims or possesses.12 visionGraph places it under the AI Research Area hub.
The capability module refines the core class into a semantic, high-level description written for natural-language understanding and fuzzy matching by language models.3 It carries a capabilityExpression string and a self-assessed capabilityConfidence between 0.0 and 1.0. It links to an Atomic Skill through hasSkill, to a Capability Constraint and to a Capability Provenance; visionGraph also relates it to Capability Advertisement.
An agent has capabilities through hasCapability, whereas an Agent Task requires them through requiresCapability. The two relations are distinct: an agent is never a task, and a capability can be required without any particular agent yet holding it. A delegation’s scope ranges over capabilities.
Two subclasses are declared in the security-binding module: Critical Capability and Restricted Capability. They are siblings and disjoint.
Footnotes
-
Semantic Agent Communication Community Group, “Core Agent Ontology” (Agent Ontology, CC BY 4.0). https://github.com/s-agent-comm/agent-ontology/blob/750e06d276dd568313a6c03c4e78203887d3fa4b/ontologies/core.ttl ↩
-
Semantic Agent Communication Community Group, “Capability Ontology” (Agent Ontology, CC BY 4.0). https://github.com/s-agent-comm/agent-ontology/blob/750e06d276dd568313a6c03c4e78203887d3fa4b/ontologies/capability.ttl ↩
-
Semantic Agent Communication Community Group, “Threat Model Ontology” (Agent Ontology, CC BY 4.0). https://github.com/s-agent-comm/agent-ontology/blob/750e06d276dd568313a6c03c4e78203887d3fa4b/ontologies/threat-model.ttl ↩