Anand Rao and Michael Georgeff: Belief, Desire and Intention as Agent Architecture
Anand Rao and Michael Georgeff connected philosophical models of belief, desire and intention to implementable architectures for rational software agents.
Rational agents needed an architecture between logic and software
By the early 1990s, researchers had rich formal theories of rational action and increasingly capable software agents, but the two traditions did not always meet. Anand Rao and Michael Georgeff’s BDI work sought to connect them by organizing an agent around three familiar mental attitudes: beliefs, desires and intentions.[1]
The aim was computational, not psychological imitation for its own sake. BDI provided a vocabulary for representing what an agent takes to be true, what outcomes it would like, and which courses of action it has committed to pursuing.
Beliefs represent the agent’s informational state
In BDI models, beliefs capture the information on which the agent acts. They need not be objectively true; they are the agent’s current model of the world. This distinction matters in dynamic environments because decisions must often be made from incomplete or outdated information.[2]
Belief is not database truth
Treating information as belief makes room for revision. An agent can update its state when new observations arrive instead of assuming that every stored proposition is permanent and globally authoritative.
Distributed agents can hold different beliefs
In multi-agent settings, two agents may observe different events and therefore act from different informational states. BDI architectures provide a structured way to reason about that divergence.
Desires represent possible objectives
Desires capture motivational states or outcomes the agent may wish to bring about. An agent can have several desires that cannot all be realized simultaneously. The architecture therefore requires deliberation rather than simply treating every goal as an immediate command.[1]
This separates the space of possible goals from the smaller set of actions an agent actually commits resources to.
Intentions stabilize action over time
Intentions are central because rational behavior cannot consist of reconsidering every possible goal at every instant. Once an agent commits to a plan, that commitment provides continuity until circumstances justify reconsideration. Rao and Georgeff’s formal work explored how such commitments can be represented and revised.[3]
Commitment solves a deliberation problem
Without some persistence, an agent can waste effort repeatedly choosing among the same alternatives. Without any reconsideration, it can stubbornly follow obsolete plans. BDI research therefore focuses on the balance between commitment and responsiveness.
The abstract architecture made the model computational
Rao and Georgeff’s 1992 work on an abstract architecture for rational agents described how BDI states could be connected to an execution cycle.[3] This helped turn philosophical categories into an implementable software architecture.
The result was a bridge between formal semantics and practical agent systems rather than a purely descriptive theory.
BDI Agents: From Theory to Practice showed application scale
Their 1995 ICMAS paper explicitly framed the challenge as integrating theoretical foundations, practical implementations and large-scale applications. It discussed air-traffic management as a domain where autonomous reasoning had to operate under dynamic, high-stakes conditions.[1]
The paper became a widely cited statement of BDI architecture precisely because it linked abstract rationality to engineering practice.
BDI became a family of systems and languages
The BDI model influenced agent programming languages, methodologies and platforms. A 1996 workshop volume, for example, includes both AgentSpeak(L) and a methodology for systems of BDI agents by David Kinny, Michael Georgeff and Anand Rao.[5]
Rao’s later publication record shows BDI work spanning formal models, verification, real-time reasoning and agent-oriented methodologies.[4]
The architecture separated reasoning concerns
Belief update, goal generation, intention selection and action execution could be studied as distinct but connected processes. That decomposition made BDI useful both for theory and for software engineering.
Why Rao and Georgeff belong in multi-agent history
BDI offered one of the clearest architectures for agents that must act repeatedly in changing environments. It supplied a durable vocabulary for informational state, motivation and commitment while also supporting executable systems.[1][2]
Modern LLM agents use very different internal machinery, but they still confront related architectural questions: what does the agent know, what is it trying to accomplish, and what plan is it currently committed to? BDI remains a major historical answer.
Works Cited
- 01
- 02AAAI — BDI Agents: From Theory to Practice PDF cdn.aaai.org
- 03APICe — An Abstract Architecture for Rational Agents rtsi16.apice.unibo.it
- 04
- 05Springer — Agents Breaking Away: MAAMAW '96 Proceedings link.springer.com
CodeHistory is a living archive. Citations document the evidence used for this edition; later evidence may refine the account.
Submit a research lead