Methodology · Agent Constructionprovenverified

BDI Agent Construction Methodology

also known as Belief-Desire-Intention agent construction, Jason/JaCaMo construction

Applies to: agentmulti-agent-system

Tags: bdibelief-desire-intentionagent-oriented-programmingjasonjacamo

Build an agent from three parts that map to how people act on purpose. The beliefs are what the agent holds true. The desires are the goals it may take on. The intentions are the goals it has committed to and is acting on. A reasoning cycle drives it: each round it picks a plan from a plan library to meet a goal. This is the belief-desire-intention (BDI) approach from the Jason and JaCaMo line of tools. Beliefs are first-class data. Plans are first-class code. The cycle is out in the open. It shines when an LLM-only agent cannot show its goals or explain its reasoning, which BDI does by default.

Methodology process overview

flowchart TD ont[Domain ontology] --> s1[Model belief base] goals[Goal catalogue] --> s2[Enumerate desires and adoption conditions] plib[Plan library input] --> s3[Author plan library] s1 --> s2 s2 --> s3 s3 --> s4[Define selection functions] s4 --> s5[Wire deliberation cycle] s5 --> s6[Integrate with environment and organisation] s6 --> out1[BDI agent program] s6 --> out2[Deliberation cycle config] s5 -.-> cycle[Perceive -> Update beliefs -> Select event -> Choose plan -> Execute intention]

Intent. Make agents whose inner state, the beliefs, desires, and intentions, is written down and easy to inspect, so the behaviour can be explained and checked instead of just emerging.

When to apply. Reach for BDI when an agent must reason openly about goals, stick with long-running commitments, coordinate with other agents under shared rules, or stand up to an audit by a regulator. It fits multi-agent setups on tools such as Jason, JaCaMo, or Jadex. Don't apply it for one-shot generation, a light tool-using chatbot, or any task where no goal lives past a single turn. The BDI machinery is overhead for stateless work. Skip it too when the team has no appetite for symbolic agent programming.

Example scenario

A research group at a logistics firm is prototyping a multi-agent yard-management system on JaCaMo. Each truck-dispatch agent must reason about goals such as unload before dock close, prioritise refrigerated cargo, and avoid blocking lanes. It must commit to those goals over minutes to hours and coordinate with other agents under a shared organisation. They tried a pure LLM agent first and dropped it, because regulators required an auditable goal-trace. Inputs: a domain ontology of yard facts (dock_open, cargo_type, lane_blocked), a goal catalogue with priorities, and a plan library drawn from dispatcher SOPs. The team modelled the beliefs with perceptual facts refreshed every two seconds and persistent facts updated on each event. They split the goals into achievement goals (unload truck T) and maintenance goals (keep aisle 3 clear). They wrote the plan library as small triggering-event plus context plus body units. They overrode the selection rules so refrigerated cargo always wins ties. The reasoning cycle ran at 0.5 Hz. Finally they attached the agents to CArtAgO artefacts for docks and lanes, and a MOISE organisation defined the dispatcher and mover roles. What they learned: because BDI agents commit to a goal, they did not drop an unload mid-task when a higher-priority goal arrived. They queued the new goal as a fresh intention instead. The auditors accepted the BDI trace as proof of why each decision was made. An LLM ReAct trace had failed that test.

Inputs

Domain ontology — The set of facts the agent can believe: facts about the world, about other agents, and about itself.
Goal catalogue — The goals the agent may take on, each with a priority and the conditions that trigger it.
Plan library — The agent's how-to knowledge: plans filed by the event and context that triggers them.

Outputs

BDI agent program — An agent definition, for example in AgentSpeak or Jason, with its beliefs, goals, and plan library.
Deliberation cycle configuration — The rules for picking which event, plan, and intention to act on, plus the settings for the reasoning cycle.

Steps (6)

Model the belief base
Decide the facts the agent holds about the world, about itself, and about other agents. Split them into perceptual facts, refreshed each cycle, and persistent facts that stay until something changes them.
Enumerate desires and adoption conditions
List the goals the agent may pursue. Give each one a trigger and a priority. Tell achievement goals (reach a state) apart from maintenance goals (keep a state true).
Author the plan library
Write each plan as a trigger, a context, and a body. The body is a list of belief updates, sub-goals, and outside actions. Plans are the unit of how-to knowledge. Keep them small and easy to combine.
usesBDI Agent Skill Library
Define selection functions
Say how the agent picks among events, plans, and intentions when more than one applies. Jason and JaCaMo give you defaults, but override them for any real domain.
Wire the deliberation cycle
Set up the perceive-decide-act cycle: read the environment, update beliefs, pick an event, choose a plan, run the intention. Tune the cycle rate to fit the domain.
usesBDI Agent
Integrate with environment and organisation
On JaCaMo, attach the agent to an environment (CArtAgO artefacts) and a social organisation (MOISE roles, groups, and schemes). This is where multi-agent coordination is grounded.
usesRole Assignment Inter-Agent Communication

Framework-specific instructions

Pick a framework and generate a framework-targeted rewrite of this methodology's steps.

Choose framework

AI-generated for Agent Development Kit (ADK) (Google) — verify against official docs.

Principles

Beliefs, desires, and intentions are real artefacts you can read, inspect, and debug.
Plans come from a library, not from improvising. Combine plans instead of hoping for emergence.
Commitment matters. An intention stays until it is met, dropped, or proven impossible.
Program the environment and the organisation alongside the agent. Do not assume them.

Known failure modes (2)

Related patterns (5)

Related compositions (2)

Related methodologies (1)

Agentic Workflow Construction★★
7 steps
Make agent authors name the four parts and the freedom level before they code, so a failure points to one part instead of smearing across a vague agent.

Sources (2)

Provenance

Added to catalog: 2026-05-24
Last updated: 2026-05-27
Verification status: verified

Methodology process overview

Steps (6)

Model the belief base

Enumerate desires and adoption conditions

Author the plan library

Define selection functions

Wire the deliberation cycle

Integrate with environment and organisation