A BDI agent, or belief-desire-intention agent, is a software agent whose behavior is organized around three mental attitudes: beliefs about the state of the world, desires describing the outcomes it would like to achieve, and intentions, the goals it has committed to pursue through concrete plans. Within AI agents, a field of artificial intelligence, the BDI model is one of the most established architectures for building agents that act deliberately in changing environments.
The approach originates in the philosopher Michael Bratman’s theory of practical reasoning, which argues that rational agents with limited resources commit to intentions instead of re-deliberating constantly. Anand Rao and Michael Georgeff turned this idea into a formal and executable model in the early 1990s, and it has since shaped agent programming languages, multi-agent simulations and industrial control systems.
The three attitudes of the BDI model
The BDI model separates what an agent knows, what it wants and what it has decided to do. Keeping these attitudes distinct allows the agent to change its mind about goals without losing track of facts, and to revise plans without abandoning its goals.
| Attitude | Represents | Typical implementation |
|---|---|---|
| Beliefs | Information about the world, possibly incomplete or wrong | Belief base of facts or logical literals |
| Desires | Candidate goals the agent could pursue | Goal set, often with priorities |
| Intentions | Goals the agent has committed to, with chosen plans | Intention stack of partially executed plans |
| Plans | Recipes for achieving goals in a given context | Plan library written by developers |
| Events | Changes in beliefs or newly adopted goals | Event queue that triggers plan selection |
Beliefs are deliberately called beliefs rather than knowledge, because they may be outdated or false. Desires may conflict with each other, whereas intentions must be mutually consistent, since the agent actually acts on them.
How the BDI reasoning cycle works
A BDI agent runs a continuous deliberation cycle that interleaves perception, reasoning and action. Each iteration is short, so the agent can react to new events while still pursuing long-term commitments.
- Perceive: read sensor input or messages and update the belief base.
- Generate options: derive which goals are relevant given current beliefs and events.
- Deliberate: filter options into intentions, respecting priorities and existing commitments.
- Select a plan: pick an applicable plan whose context condition holds under current beliefs.
- Execute one step: perform an action or post a subgoal, then return to perception.
A defining feature is commitment. Once an intention is adopted, the agent keeps it until it succeeds, becomes impossible or is no longer wanted. Rao and Georgeff described different commitment strategies, from blind commitment to single-minded and open-minded commitment, which control how readily the agent drops intentions.
| Strategy | Drops an intention when | Effect on behavior |
|---|---|---|
| Blind commitment | It is believed achieved | Persistent but may pursue hopeless goals |
| Single-minded commitment | Achieved or believed impossible | Balanced persistence and realism |
| Open-minded commitment | Achieved or no longer desired | Flexible but can switch goals often |
BDI compared with other agent architectures
The BDI architecture sits between purely reactive agents and agents that plan everything from first principles. It relies on precompiled plan libraries instead of searching for plans at runtime, which makes it fast and predictable but limits it to situations the designers anticipated.
| Agent type | Decision basis | Main strength | Main weakness |
|---|---|---|---|
| Simple reflex agent | Condition-action rules | Very fast reactions | No goals or memory |
| Goal-based agent | Search toward explicit goals | Finds novel solutions | Expensive planning |
| Utility-based agent | Maximizing expected utility | Handles trade-offs | Needs a utility model |
| BDI agent | Beliefs, goals and plan libraries | Responsive yet goal-directed | Limited to authored plans |
| LLM-based agent | Language model reasoning in an agent loop | Broad open-ended behavior | Less predictable, harder to verify |
Modern LLM-based agents share several BDI ideas without naming them: an agent memory or scratchpad acts as a belief base, task decomposition resembles goal adoption, and plan-and-execute agents commit to intermediate plans. Researchers increasingly combine both approaches, using a language model to propose plans while a BDI layer enforces commitments and explicit context conditions.
Implementations and practical use
The BDI model is implemented in dedicated agent programming languages and frameworks. The Procedural Reasoning System (PRS) was an early implementation used for fault diagnosis in spacecraft systems. AgentSpeak and its interpreter Jason, JACK, Jadex and 2APL followed, and they remain common in research and teaching.
Typical applications are domains where goals persist while conditions change: air traffic management and other decision-support systems, unmanned vehicles, simulations of human behavior in emergencies and training, and robotic missions. In these settings the explicit plans of a BDI agent make it easier to explain and certify than opaque learned policies.
Limitations and common misconceptions
The BDI approach has well-known weaknesses that matter when choosing it for a project. The most important is that behavior quality depends on the plan library: if no plan matches the current context, the agent simply cannot proceed.
- Not a learning architecture by default: classic BDI agents do not improve plans from experience; learning must be added separately.
- Mental terms are metaphors: beliefs and desires are data structures, not evidence of understanding or consciousness.
- Plan authoring is costly: large domains require many hand-written plans and careful context conditions.
- Formal guarantees are partial: BDI logics are expressive, but verifying a full implementation remains difficult.
The idea in brief
A BDI agent is a program designed to act a little like a sensible person planning a day. It keeps a picture of what it thinks is true, a list of things it would like to achieve, and a short list of things it has actually decided to do.
When something changes, it updates its picture and checks whether its plans still make sense. It sticks to its decisions unless there is a good reason to drop them, which keeps it from changing course every second.
A practical scenario
A logistics company uses BDI agents to control a fleet of 30 warehouse robots. Each robot believes where shelves, chargers and other robots are, desires to complete assigned picking tasks and keep its battery above 20 percent, and holds intentions such as fetching a specific pallet. Its plan library contains several dozen plans for navigation, picking, recharging and handling blocked aisles.
When an aisle is suddenly blocked, the belief update triggers an alternative routing plan while the intention to deliver the pallet stays in place. When the battery falls below the threshold, a higher-priority recharge intention suspends the delivery and resumes it afterward. Because every decision maps to a named plan, engineers can trace incidents quickly, and the fleet keeps working through disruptions without constant human intervention.