Multi-Agent System Ontology

Agent TYPE (not execution instance). Represents the schema/template discovered from the execution trace. Example: trace name 'itinerary' → Agent type 'ItineraryAgent'. Execution instances are AgentCall.

Agent execution INSTANCE extracted from the execution trace. Linked to Agent TYPE via :executesAgent. Example: trace unit 'itinerary_20250113_143052' → AgentCall_001 :executesAgent ItineraryAgent (type).

Common parent for structural capability types an Agent can use: LLM, Tool, Processing. Placeholder for attributes shared across all three.

Common parent for execution instances of a Capability: LLMCall, ToolCall, ProcessingCall. Placeholder for attributes shared across all three.

Abstract root for every node in this ontology: structural types, execution instances, and trajectory nodes alike.

Execution instances (not types) extracted from the execution trace. 1 trace unit = 1 *Call instance. These are runtime occurrences: AgentCall_001, LLMCall_gpt4_042, etc.

LLM type discovered from the execution trace

LLM call extracted from the execution trace (input prompt + output completion)

MAS instance discovered from the execution trace (which agents were composed together)

MAS execution extracted from the execution trace (orchestration of multiple agents)

Processing type discovered from the execution trace. Represents internal computation capabilities within agents.

Processing code execution inside an agent (internal computation, data transformation, validation logic, etc.). Distinct from LLMCall (external model inference) and ToolCall (external API).

Session execution extracted from the execution trace (top-level execution context)

Data state at a point in execution (prompt/input/output). Example: State_prompt_001 with content 'Book a flight to Paris'. States are connected via Transitions.

Schema/type definitions representing the static structure of a MAS. These are TYPES not instances: Agent (not AgentCall), etc.

Tool type discovered from the execution trace

Tool call extracted from the execution trace (external API/DB call)

Parent class for State/Transition — the normalized (raw unstructured text) view of a session's trajectory.

State transition representing a processing step between states. Links normalized data flow to execution instances via :representsExecution. Example: Transition_001 :inputTo State_prompt, :leadsTo State_response, :representsExecution AgentCall_itinerary_001.

MAS is composed of Agents
Domain: Agent
Range: MAS

AgentCall belongs to MASCall link.
Domain: AgentCall
Range: MASCall

Cached input tokens reused.
Domain: LLMCall
Range: integer

Number of tokens in the completion.
Domain: LLMCall
Range: integer

The raw textual content of the state (the delta produced by each call).
Domain: State
Range: string

ExecutionElement instance corresponds to this Transition (AgentCall_001 → Transition between its input/output states)
Domain: ExecutionElement
Range: Transition

Whether this structural type was explicitly declared (e.g. in an agent manifest or skill registry) rather than dynamically discovered from the execution trace. False (or absent) means inferred.
Domain: StructuralElement
Range: boolean

Human-readable description of this element.
Domain: nc7e8507f68a948789704fe2a9153f0d5b2
Range: string

Duration in milliseconds.
Domain: ExecutionElement
Range: double

End time of the execution as a Unix timestamp (seconds, high precision).
Domain: ExecutionElement
Range: decimal

Generic link from an execution instance to the structural type it is an instance of. See subproperties :executesMAS, :executesAgent, :executesLLM, :executesTool, :executesProcessing for the specific pair.
Domain: ExecutionElement
Range: StructuralElement

AgentCall instance executes this Agent type (e.g., AgentCall_001 executes ItineraryAgent)
Domain: AgentCall
Range: Agent

LLMCall instance is of this LLM structural type (e.g., GPT-4, Claude).
Domain: LLMCall
Range: LLM

MASCall instance executes this MAS type
Domain: MASCall
Range: MAS

ProcessingCall executes this Processing structural type (e.g., DataNormalizer, ResponseFormatter).
Domain: ProcessingCall
Range: Processing

Session executes this MAS configuration
Domain: Session
Range: MAS

ToolCall instance is of this Tool structural type (e.g., WeatherAPI).
Domain: ToolCall
Range: Tool

Reason generation stopped. Values: 'stop', 'length', 'content_filter', 'tool_calls'.
Domain: LLMCall
Range: string

MASCall orchestrates AgentCall executions (1:N — one MAS call coordinates multiple agent calls)
Domain: MASCall
Range: AgentCall

Points to the final/terminal State of an execution. Subproperty of :hasState for easier querying.
Domain: ExecutionElement
Range: State

Points to the initial/starting State of an execution. Subproperty of :hasState for easier querying.
Domain: ExecutionElement
Range: State

Legacy shortcut (denormalized) from AgentCall to LLMCall. Canonical typing links are mas:executesLLM on LLMCall and containment topology edges. Consumers should avoid emitting this edge in normalized KG outputs.
Domain: AgentCall
Range: LLMCall

Session contains MASCall executions (1:N — one session can have multiple MAS calls, e.g. retries/continuations).
Domain: Session
Range: MASCall

Denormalized shortcut from AgentCall or MASCall to a ProcessingCall it directly contains. Canonical typing link is mas:executesProcessing on Processing.
Domain: nc7e8507f68a948789704fe2a9153f0d5b6
Range: ProcessingCall

ExecutionElement has associated States in its trajectory. Example: AgentCall_001 :hasState State_reasoning
Domain: ExecutionElement
Range: State

Legacy shortcut (denormalized) from AgentCall to ToolCall. Canonical typing links are mas:executesTool on ToolCall and containment topology edges. Consumers should avoid emitting this edge in normalized KG outputs.
Domain: AgentCall
Range: ToolCall

Generic identifier for this element. Subclasses may declare a more specific, formally-formatted identity/MERGE key on top of this (e.g. :callId, :stateNodeId, :transitionId, :sessionId) — those are not replacements for :id, they are stricter identity contracts for specific node kinds.
Domain: Element
Range: string

Source State feeds into this Transition (State → Transition). Declared owl:inverseOf :leadsTo as a schema-level pairing (this codebase does no OWL inference, so this does not imply automatic mirrored triples) — the two are independently-populated adjacent hops used together to walk a trajectory's State → Transition → State chain.
Domain: State
Range: Transition

Transition leads to its next/target State (Transition → State). Also serves as the canonical 'target state of this Transition' property for endpoint cardinality checks (see mas-shapes.ttl TransitionEndpointsShape).
Domain: Transition
Range: State

Human-readable name of this element.
Domain: Element
Range: string

ID of the corresponding parent span in OTel traces, where one exists.
Domain: ExecutionElement
Range: string

Number of tokens in the prompt.
Domain: LLMCall
Range: integer

LLM provider name.
Domain: LLM
Range: string

Canonical execution edge: Transition -> ExecutionElement. This object property is the normalized topology link.
Domain: Transition
Range: ExecutionElement

Unique response ID from the LLM provider. Used for provider-side debugging.
Domain: LLMCall
Range: string

Unique session identifier. Primary key for Session nodes (MERGE key). Carried as a provenance attribute on State and Transition nodes to make their scope explicit without a join.
Domain: ExecutionElement
Range: string

ID of the corresponding span in OTel traces, where one exists.
Domain: ExecutionElement
Range: string

Start time of the execution as a Unix timestamp (seconds, high precision).
Domain: ExecutionElement
Range: decimal

Model temperature setting (sampling randomness, typically 0.0-2.0).
Domain: LLMCall
Range: double

Total tokens used (prompt + completion).
Domain: LLMCall
Range: integer

Generic link from an Agent type to a structural capability it can use. See subproperties :usesLLM, :usesTool, :usesProcessing for the specific capability kind.
Domain: Agent
Range: Capability

Agent type can use this LLM type.
Domain: Agent
Range: LLM

Agent type can use this Processing type.
Domain: Agent
Range: Processing

Agent type can use this Tool type.
Domain: Agent
Range: Tool