BDI-O: Cognitive Agent Ontology

• BDI-O is a formally structured and modular ontology that encodes beliefs, desires, and intentions for rational agents using OWL, DL, and ODPs.
• It operationalizes mental state transitions through defined processes and temporal logic, ensuring dynamic, explainable, and consistent agent reasoning.
• BDI-O underpins practical applications in multi-agent systems and neuro-symbolic AI by enabling semantic interoperability and robust coordination.

The Belief-Desire-Intention Ontology (BDI-O) is a formally structured, modular ontology for representing the cognitive architecture of rational agents in artificial intelligence, cognitive science, and multi-agent systems. It instantiates the Belief-Desire-Intention (BDI) model with Description Logic (DL), Web Ontology Language (OWL), and Ontology Design Patterns (ODPs), supporting dynamic mental state transitions, alignment with foundational ontologies, and operational integration with neuro-symbolic and web-scale systems. BDI-O provides an interoperable, semantically precise foundation for encoding mental states, their interrelations, and their evolution through processes, enabling advanced reasoning, explainability, and cross-agent compatibility within the Web of Data (Zuppiroli et al., 21 Nov 2025).

1. Conceptual Motivation and Foundational Architecture

BDI-O formalizes the BDI model—originating with Bratman’s theory of intention—by capturing the three central cognitive attitudes necessary for rational agency: beliefs (informational stance), desires (motivational stance), and intentions (deliberative commitments). Prior to BDI-O, most BDI implementations lacked a rigorous, interoperable, and modular ontological formalism, limiting their integration into Semantic Web, neuro-symbolic, and multi-agent ecosystems.

Key principles driving BDI-O’s architecture include:

• Explicit encoding of beliefs, desires, intentions, mental processes, plans, goals, and their mutual relationships as OWL classes and object properties.
• Division of mental entities into endurants (mental states: Belief, Desire, Intention) and perdurants (mental processes: BeliefProcess, DesireProcess, IntentionProcess), leveraging DOLCE UltraLite (DUL) patterns for events and situations.
• Support for dynamic, time-indexed transitions between mental states, enabling representation of agent deliberation, intention reconsideration, plan adoption, and explanation.

BDI-O is published as a modular ODP and aligns systematically with foundational ontologies, maximizing reuse and interoperability (Zuppiroli et al., 21 Nov 2025).

2. Ontological Structure and Modular Patterning

BDI-O defines a comprehensive core comprising classes, properties, and axioms. Below is a representative abstract of the OWL/Turtle schema:

Class	Superclass	Description
:Agent	dul:Agent	Agentive entity
:MentalEntity	d0:CognitiveEntity	All mental phenomena
:MentalState	:MentalEntity	State: Belief, Desire, Intention
:Belief	:MentalState	Informational mental state
:Desire	:MentalState	Motivational state
:Intention	:MentalState	Deliberative commitment
:MentalProcess	d0:Activity, :MentalEntity	Dynamic process affecting states
:Plan	dul:Plan	Structured sequence of actions
:Goal	dul:Goal	Target world state
:PlanExecution	dul:PlanExecution	Realization of a plan

Representative object properties:

• bdi:perceives (Agent → WorldState), bdi:hasBelief (Agent → Belief), bdi:motivates (Belief → Desire), bdi:fulfills (Intention → Desire), bdi:specifies (Intention → Plan), bdi:bringsAbout (PlanExecution → WorldState), bdi:atTime (MentalEntity → TemporalEntity).

DL axioms include, e.g., $Belief \sqsubseteq \exists motivates.Desire$ and $Intention \sqsubseteq \exists fulfills.Desire$, capturing the generative and fulfillment relationships among mental states. Temporal anchoring is enforced: $$MentalEntity \sqsubseteq \exists atTime.TemporalEntity$$ (Zuppiroli et al., 21 Nov 2025).

The ontology is organized into modular patterns (EventCore, Situation, Provenance, BasicPlan, TimeIndexedSituation) mapped to specific sets of competency questions and validated with SPARQL queries. Explicit alignments to DUL, via bdi-dul.owl, guarantee compatibility with common upper ontologies.

3. Dynamic Transitions, Reasoning, and Constraints

BDI-O explicitly represents transitions among mental states through formally defined processes:

• BeliefProcess: $BeliefProcess \sqsubseteq \forall generates.Desire$
• DesireProcess: $$DesireProcess \sqsubseteq \forall generates.Intention$$
• IntentionProcess: $$IntentionProcess \sqsubseteq \forall generates.Intention$$, with $Intention \sqsubseteq \exists specifies.Plan$

These transitions may be encoded as production rules or higher-order temporal logic statements:

• If an agent has a Belief $B$ which motivates a Desire $D$, then a BeliefProcess generates $D$.
• If an agent has a Desire $D$, a DesireProcess generates an Intention $I$ (Zuppiroli et al., 21 Nov 2025).

For domain integrity, SHACL shapes enforce constraints such as: every Belief instance must refer to precisely one WorldState and possess at least one Justification.

4. Integration with LLMs and Logic-Augmented Generation

BDI-O is engineered to support Logic-Augmented Generation (LAG) workflows with LLMs:

• An OWL/Turtle serialization of the relevant BDI-O subgraph, alongside competency questions and RDF templates, frames the LLM prompt.
• The system-generated RDF triples are reasoned over for ontological consistency and mapped back to corresponding agent mental-state structures.

This enables LLM outputs to be grounded in formally specified agent models, improving inferential coherence and ensuring logical consistency. In experimental evaluation with GPT-4, ontology-augmented models detected +22% more inconsistencies and achieved 100% coverage on specified modelling tasks compared to standard baselines (Zuppiroli et al., 21 Nov 2025).

The triple↔state mapping is bidirectional: input RDF triples are mapped onto in-memory Belief, Desire, or Intention objects, while agent state updates are serialized back into Turtle for external consumption.

5. Operationalization: The Semas T2B2T Paradigm

The Semas platform operationalizes BDI-O under the Triples-to-Beliefs-to-Triples (T2B2T) paradigm:

• Streaming RDF triples (e.g., from IoT sources) are ingested and matched against BDI-O patterns.
• Detected WorldState triples trigger creation of Beliefs via production rules; subsequent processes generate Desires, Intentions, and PlanExecutions, all materialized as RDF output.
• The reasoning engine is Prolog-style, enabling specification and execution of rules of the form $[HEAD] / [CONDITIONALS] \gg [TAIL]$ linking semantic conditions to BDI-O mental-state transitions.
• Consistency with BDI-O axioms is preserved throughout round-trip serialization; scalability has been empirically demonstrated at $>$200 inferences/sec on knowledge bases of $\sim$5K triples (Zuppiroli et al., 21 Nov 2025).

SPARQL and Prolog queries are directly usable for extracting agent mental states at specific time instants, facilitating explainable, time-aware multi-agent reasoning.

6. Selected Use Cases, Evaluation, and Expressivity

Applications to high-agreement tasks (e.g. MS-LaTTE) confirmed BDI-O’s capability to:

• Detect contradiction in agent intentions infeasible under current beliefs and world states.
• Model complex agent deliberation chains, capturing process triggers, time stamps, and rationale links.

In implementation, all generated RDF structures conform to the ontology’s axioms, and the modular patterning supports extensibility to domain-specific agency.

BDI-O’s design has enabled the representation of intention–intention conflicts (with further explicit axiomatisation in future work), and provides a foundation for reasoning about shared mental states, joint intentions, and negotiation in multi-agent environments.

7. Future Directions and Theoretical Extensions

Planned enhancements to BDI-O include:

• Formalization of intention–intention conflict detection and resolution, e.g., through an IncompatibleWith property and associated conflict-resolution processes.
• Integration of metareasoning patterns (e.g., Russell & Wefald’s ODPs) for managing resource-bounded deliberation and intention reconsideration.
• Extension of temporal reasoning with more expressive modal logics (such as CTL or SWRL-based branching time).
• Multi-agent coordination capabilities, support for jointly held beliefs and intentions, and richer negotiation behaviors.
• Publication of BDI-O instances on DBpedia, Wikidata, and registration in vocabularies such as LOV for Web of Data interoperability, along with alignment to provenance standards such as PROV-O (Zuppiroli et al., 21 Nov 2025).

A plausible implication is that BDI-O, due to its modular, ontologically grounded, and operationally testable design, forms a central semantic infrastructure for explainable neuro-symbolic and web-scale multi-agent systems.

8. Logical Foundations: Dynamic Preference Logic and Priority Graphs

The dynamic and agentive aspects of BDI-O can be further analyzed through the lens of Dynamic Preference Logic (DPL), which encodes BDI attitudes in terms of preference orders, modalities, and their evolution over agent programs (Souza et al., 2019):

• Preference models $M = \langle W, \leq_P, \leq_D, \nu \rangle$, with two preorders (plausibility for belief, desirability for desire), and interpretations of conditional belief and desire as modalities.
• Rational agent programs specified as tuples $\langle K, G_P, G_D, I \rangle$ with priority graphs $G_P$, $G_D$ reflecting doxastic and desiderative rankings; all model dynamics are realized as transformations on these priority graphs.
• Rigorous formal treatment ensures completeness: every agent model corresponds to some finite priority graph; deliberative updates (belief revision, intention reconsideration, plan adoption) correspond to well-defined syntactic manipulations, ensuring no semantic gap between agent-level deliberation and underlying ontological representation.

This formal underpinning situates BDI-O as both a semantic and algorithmic bridge between agent programming, dynamic epistemic logic, and ontology-driven AI (Souza et al., 2019).

---

References

• "The Belief-Desire-Intention Ontology for modelling mental reality and agency" (Zuppiroli et al., 21 Nov 2025)
• "A Dynamic Preference Logic for reasoning about Agent Programming" (Souza et al., 2019)