Bdi Agency Model

SKILL.md: The Belief-Desire-Intention Model of Agency

license: Apache-2.0

Metadata

• Skill Name: bdi-agency-model
• Source: "The Belief-Desire-Intention Model of Agency" (Georgeff, Pell, Pollack, Tambe, Wooldridge)
• Domain: Intelligent systems architecture, multi-agent systems, adaptive planning
• Activation Triggers:
  • Designing autonomous agents or adaptive systems
  • Debugging planning/replanning behavior
  • Evaluating when systems should reconsider decisions
  • Architecting systems for dynamic environments
  • Coordinating multiple intelligent agents
  • Questions about goal vs. task representation

When to Use This Skill

Load this skill when facing:

• Architecture decisions: How to structure an intelligent agent's internal representation and control flow
• Commitment problems: System thrashing between plans or stubbornly pursuing obsolete goals
• Dynamic environment design: Systems must adapt to changing conditions without complete replanning
• Resource constraints: Limited computation time for decision-making in real-time systems
• Recovery challenges: Systems can't automatically recover when plans fail
• Multi-agent coordination: Multiple agents need to form and maintain joint commitments
• Goal vs. implementation gap: Distinguishing what systems should achieve from how they achieve it
• Learning integration: Adding adaptive capabilities to planning systems

This skill applies beyond robotics/AI agents—it's relevant for any system that must make decisions over time with partial information, including business process systems, game AI, automated trading, and organizational design.

Core Mental Models

1. The Three Components Are Computationally Necessary, Not Optional

Beliefs (informational state), Desires (motivational state), and Intentions (deliberative state) aren't philosophical conveniences—they're architectural necessities for resource-bounded agents in dynamic environments.

• Beliefs: What you currently know about the world (may be wrong, incomplete)
• Desires: All goals/outcomes you'd like to achieve (may be conflicting, unachievable)
• Intentions: Subset of desires you've committed to pursuing (chosen, planned, actionable)

Why separate? Because conflating them creates impossible computational requirements:

• Beliefs + Desires merged → must constantly recompute feasibility of every desire
• Desires + Intentions merged → no distinction between "would be nice" and "actively pursuing"
• Beliefs + Intentions merged → when beliefs change, must immediately replan everything

The separation enables local updates and selective reasoning.

2. The Commitment-Reconsideration Balance

The core tension:

• Replan too often (classical decision theory) → waste resources thrashing, never execute
• Replan too rarely (conventional software) → miss opportunities, pursue obsolete goals

The empirical finding (bold vs. cautious agents):

• Static environments → bold agents win (commit strongly, rarely reconsider)
• Highly dynamic environments → cautious agents win (reconsider frequently)
• Real environments → need meta-level strategy for when to reconsider

Rational commitment = maintaining intentions until there's sufficient reason to reconsider, where "sufficient" depends on environment dynamics and computational costs.

3. Task-Orientation vs. Goal-Orientation

Task-oriented systems: Execute procedures without maintaining why

• "Run script X" → executes blindly
• Cannot automatically recover from failures
• Cannot exploit opportunistic improvements
• Cannot explain behavior

Goal-oriented systems: Maintain explicit desired states and plans to achieve them

• "Achieve state Y via plan Z" → knows current beliefs, goal, and chosen approach
• Can detect when goals already achieved (stop early)
• Can recognize when goals become impossible (abandon gracefully)
• Can find alternative plans when current fails
• Can explain actions in terms of goals

The gap: Most software is task-oriented. BDI provides the minimal architecture for goal-orientation.

4. Intentions as Computational Filters

Intentions don't just represent commitments—they actively constrain future reasoning:

• Filter practical reasoning: Only consider options compatible with existing intentions
• Reduce option space: Avoid reconsidering entire action space at each step
• Enable predictability: Others can predict your behavior (crucial for coordination)
• Structure persistence: Provide default continuity unless reasons to change

This is not bounded rationality or satisficing—it's a rational strategy for resource-bounded agents. The alternative (consider all options always) is computationally intractable.

5. Convergent Evolution Signals Universal Principles

BDI (philosophy/logic tradition) and Soar (cognitive psychology tradition) independently converged on isomorphic architectures:

Implication: These components aren't arbitrary design choices but discovered necessities—like multiple cultures independently inventing the wheel. Any system facing similar constraints (bounded resources, partial information, dynamic environments, real-time action) will need similar structures.

Decision Frameworks

When to Reconsider Intentions

IF environment is highly static THEN bias toward bold commitment (rarely reconsider) IF environment is highly dynamic THEN bias toward cautious reconsideration (frequently reevaluate)

IF computational resources abundant THEN can afford more frequent reconsideration IF computational resources scarce THEN must commit more strongly to amortize planning cost

IF action execution is expensive THEN reconsider before acting IF deliberation is expensive THEN commit and act quickly

IF new information directly contradicts intention's preconditions THEN reconsider immediately IF new information merely opens alternatives THEN reconsider only if significantly better

When to Separate Beliefs, Desires, and Intentions

IF system must act in real-time THEN needs intentions (can't replan from desires every cycle) IF system faces conflicting goals THEN needs desire/intention distinction (desires can conflict, intentions cannot) IF system must coordinate with others THEN needs explicit intentions (for commitment communication) IF system must explain behavior THEN needs all three (explain actions via intentions, intentions via desires, desires via beliefs) IF system must adapt to failures THEN needs goal-oriented representation (to find alternative means)

Architecture Selection

IF environment is static AND goals are clear AND plans rarely fail THEN task-oriented architecture may suffice IF environment changes OR goals conflict OR plans frequently fail THEN need BDI-style architecture IF multiple agents must coordinate THEN need explicit commitment mechanisms (load multi-agent reference) IF system must learn from experience THEN need BDI + learning extensions (load learning reference)

Reference Table

Anti-Patterns

1. Replanning Everything Always

Symptom: System constantly reconsiders all options, thrashes between plans, never completes actions Root cause: Treating every belief update as requiring complete replanning (classical decision theory mistake) BDI insight: Rational commitment means maintaining intentions despite minor belief changes

2. Blind Task Execution

Symptom: System continues executing plan after goal achieved or becomes impossible Root cause: Task-oriented architecture without goal representation BDI insight: Maintain explicit relationship between intentions, plans, and desires to enable goal-aware execution

3. Desire-Intention Conflation

Symptom: System tries to pursue conflicting goals simultaneously, resources spread too thin Root cause: Not distinguishing between "would like to achieve" and "committed to pursuing" BDI insight: Desires can conflict; intentions cannot (must resolve conflicts during deliberation)

4. Intention Without Commitment

Symptom: System abandons plans at first difficulty; erratic, unpredictable behavior Root cause: Treating intentions as weak preferences rather than commitments BDI insight: Intentions constrain future reasoning—they're not just goals but committed goals

5. Static Plans in Dynamic Environments

Symptom: System fails when environment changes; no recovery from failures Root cause: Conventional software approach (plan = fixed procedure) BDI insight: Plans must be monitored against goals; reconsideration needed when environment dynamics high

6. No Meta-Level Control

Symptom: Fixed reconsideration strategy fails across different environment dynamics Root cause: Not adapting commitment strategy to environment characteristics BDI insight: Need meta-level reasoning about when to reconsider (the central open problem)

7. Single-Agent BDI in Multi-Agent Settings

Symptom: Coordination failures; agents undermine each other's plans Root cause: Using individual BDI without social commitment mechanisms BDI insight: Multi-agent requires joint intentions and commitment communication protocols

Shibboleths: How to Recognize Deep Understanding

Surface-Level (Just Read Summary)

• "BDI means beliefs, desires, and intentions"
• "It's based on Bratman's philosophy"
• "You need to balance commitment and flexibility"

Intermediate (Understood Concepts)

• Can explain why three components are computationally necessary, not just what they are
• Recognizes commitment-reconsideration as empirical trade-off with environment-dependent optima
• Distinguishes task-oriented vs. goal-oriented architectures by recovery capabilities

Deep (Internalized)

• Sees BDI as resource management strategy: Recognizes that intentions exist to constrain reasoning space, not just represent commitments
• Treats "when to reconsider" as the central problem: Understands bold/cautious experiments show no universal answer—context determines strategy
• Recognizes convergent evolution: Points to Soar-BDI mapping as evidence these are discovered necessities, not designed choices
• Distinguishes individual vs. social commitments: Knows single-agent BDI is insufficient for multi-agent coordination
• Identifies learning integration as fundamental gap: Acknowledges BDI provides structure but not adaptation mechanisms

Master (Can Apply Cross-Domain)

• Recognizes BDI patterns in non-AI systems (business processes, organizational design, personal productivity)
• Can diagnose "commitment pathologies" (thrashing, stubbornness) and prescribe architectural remedies
• Designs meta-level control strategies tuned to specific environment dynamics
• Integrates learning mechanisms while preserving commitment semantics
• Uses task/goal distinction to evaluate system explainability and recovery capabilities

Key diagnostic question: "Why can't you just merge desires and intentions?"

• Shallow answer: "Because you need to distinguish what you want from what you're doing"
• Deep answer: "Because desires can be conflicting and unchosen—deliberation is the computational process of resolving conflicts and resource constraints to select a feasible subset as intentions. Merging them eliminates the structure that makes deliberation tractable and commitments meaningful."