Belief–Desire–Intention Software Model¶
A BDI software model organizes an agent's information, candidate objectives, and persisting intentions through an interpreter that selects and revises action as events occur.
Core Idea¶
A belief–desire–intention (BDI) software model organizes an acting software agent around its revisable information (beliefs), possible objectives (desires), and selected courses that persist across decisions (intentions). An interpreter uses these roles to consider options, act, and reconsider commitments when events change the situation. It is the relation among the roles that matters; a program with three matching variable names is not enough.[ref-efd6fb235c32][ref-b958a0efe2ad]
A desire need not be stored in a dedicated data structure: possible goals can be computed when needed. An intention is more than a currently executing step because it constrains later deliberation until a policy releases it. Practical implementations often use triggered plans and intention stacks, but those are implementation choices rather than requirements of every BDI specification.[ref-efd6fb235c32][ref-b958a0efe2ad]
Scope of Application¶
The Procedural Reasoning System (PRS) used beliefs, goals, conditional knowledge areas, an interpreter, and a stack of active processes. Its Flakey robot was tested in research route navigation and interruption for a keyboard-simulated jet fault. Physical tool retrieval was an envisaged task, not a demonstrated test result.[^ref-b958a0efe2ad]
OASIS supplied a different agent setting for air-traffic-management research. Aircraft and global agents handled arrival and wind information; a sequencer retained a schedule until it believed the aircraft had landed in sequence or no longer believed the next aircraft could meet its assigned arrival. The 1995 account describes parallel evaluation receiving live radar, not authoritative operational control.[^ref-efd6fb235c32]
Clarity¶
Possible goals, adopted goals, and persisting intentions should not be collapsed. The distinction tells us which objectives the agent may consider and which course it has committed to pursue. Nor must beliefs, desires, and intentions be three stored tables: Rao and Georgeff explicitly allow desires to be generated on demand. An operating interpreter, rather than the labels alone, makes the software architecture.[^ref-efd6fb235c32]
The existing Belief–Desire–Intention Model describes the higher-level practical-reasoning account that inspired this software architecture. A formal specification of attitudes can likewise describe BDI without implementing an acting agent. These are related, distinct objects.[^ref-efd6fb235c32]
Manages Complexity¶
Agent behavior involves facts, competing objectives, events, procedures, and unfinished courses. Five questions compress that detail: what does the agent believe, what could it pursue, what has it committed to, which responses are available, and when should it act or reconsider? Wrong information calls for belief revision; competing possible goals call for selection; an infeasible intention calls for reconsideration. The map helps identify which part of an agent design needs repair.[ref-efd6fb235c32][ref-b958a0efe2ad]
Abstract Reasoning¶
Trace a new event through a candidate agent. Does it change beliefs, produce or remove possible goals, alter the available plans, trigger deliberation, or change an adopted intention? Then ask whether the agent acts and revises its state. If intention state never affects later choice, the BDI label has not established the distinctive architecture.[^ref-efd6fb235c32]
A schedule that the OASIS sequencer no longer believes feasible gives a concrete reason to reconsider. PRS interruption for a simulated urgent fault gives another. Neither example implies that an agent should deliberate after every event or persist through every disruption; the commitment policy must balance decision time and responsiveness.[ref-efd6fb235c32][ref-b958a0efe2ad]
Knowledge Transfer¶
BDI organization can transfer literally between robot control and other software-agent settings when represented information, candidate objectives, persisting intentions, and an event-responsive interpreter remain operative. Their sensors, procedures, and validation standards must still be designed for each setting. Using the terms for a human team is analogy to this software model, even though the model drew on a philosophical account of intention.[ref-efd6fb235c32][ref-b958a0efe2ad]
The strict parent here is Software Architecture: BDI is a particular way to organize software-agent elements and responsibilities. The possibility of a wider information–objective–commitment pattern outside software remains a future Prime question; these two research cases do not establish it.
Example¶
PRS Flakey. The agent's belief database represented current information. Navigation and response to a simulated fault supplied candidate tasks. Active knowledge areas on a process stack supplied intention continuity and could be interrupted for higher-priority work. Conditions on beliefs and goals invoked procedural options; the interpreter executed and revised steps as events arose. Partial hierarchical knowledge areas were the practical means. The observed test was research navigation and simulated interruption, not a real space-station emergency.[^ref-b958a0efe2ad]
OASIS parallel trial. Aircraft and wind information informed beliefs; arrival times and landing order supplied candidate objectives. The sequencer's schedule intention persisted under its belief-based completion and feasibility rules. Written plan options and an event-responsive interpreter supplied option selection, execution, and revision. The account establishes parallel evaluation with live radar, not authoritative live air-traffic control.[^ref-efd6fb235c32]
Relationships to Other Abstractions¶
Current abstraction Belief–Desire–Intention Software Model Domain-specific
Parents (1) — more general patterns this builds on
-
Belief–Desire–Intention Software Model is a kind of Software Architecture Domain-specific
A BDI software model is a specific organization of software-agent state, choice, commitment, and execution.
Hierarchy path (1) — routes to 1 parentless root
- Belief–Desire–Intention Software Model → Software Architecture
Neighborhood in Abstraction Space¶
Belief–Desire–Intention Software Model sits in a sparse region of the domain-specific corpus (94th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.
Family — Unclustered & Miscellaneous (2551 abstractions)
Nearest neighbors
- Reinforcement learning — 0.82
- Inferential Theory of Learning — 0.80
- Simulation — 0.78
- Use Case — 0.77
- Frege's Puzzles — 0.77
Computed from structural-signature embeddings · 2026-10-08
Not to Be Confused With¶
Philosophical BDI model concerns practical reasoning rather than an operating software agent. Reactive rules may map events directly to actions without persisting intentions. A fixed workflow may run steps without goal selection or context-sensitive reconsideration. A particular BDI interpreter such as Jason is one implementation, not the definition of the whole class. A PRS plan library does not make fully prewritten plans universal, and PRS can interleave partial plan formation with execution.[ref-efd6fb235c32][ref-b958a0efe2ad][^ref-5a5693092351]
References¶
[^ref-efd6fb235c32]: Anand S. Rao and Michael P. Georgeff, “BDI Agents From Theory to Practice”, Proceedings of the First International Conference on Multiagent Systems (1995), printed pp. 312–319, especially pp. 313–318. The source title page prints “BDI Agents: From Theory to Practice”; the linked transcription omits the colon so the reference binder retains the complete work identity. Primary full text for the abstract and practical BDI interpreters, commitment policies, and OASIS parallel evaluation.
[^ref-b958a0efe2ad]: Michael P. Georgeff and Amy L. Lansky, “Reactive Reasoning and Planning”, Proceedings of the Sixth National Conference on Artificial Intelligence (1987), printed pp. 677–682, especially §§3.1–3.4 and §5. Primary full text for PRS structure, interleaved partial planning and execution, and the reported Flakey research test.
[^ref-5a5693092351]: Jason project, “Jason” official documentation, “About Jason” and “Jason Agent Language.” Primary project description of an AgentSpeak-based BDI interpreter; cited only for an implementation variant, not comparative efficacy.