curiositech/rao-georgeff-1995-bdi-agents-from-theory-to-practice

SKILL: BDI Agents - Practical Rational Agency Under Resource Bounds

Skill ID: bdi-agents-theory-practice
Domain: Agent architectures, AI system design, decision-making under resource constraints
Source: Rao & Georgeff, "BDI Agents: From Theory to Practice"

Decision Points

Primary Decision Tree: Commitment Strategy Selection

INPUT: Environment volatility + Goal clarity + Computation budget

IF Environment volatility = LOW (changes slower than plan execution):
  └─ IF Goal clarity = HIGH (specific, measurable outcomes):
     └─ CHOOSE: Blind commitment
        • Perceptual cue: Plan steps complete in predictable timeframes
        • Trigger reconsideration: Only when goal achieved/failed
  └─ IF Goal clarity = LOW (abstract, evolving objectives):
     └─ CHOOSE: Single-minded commitment
        • Perceptual cue: Subgoal failures indicate impossibility

IF Environment volatility = MEDIUM (changes comparable to plan duration):
  └─ IF Computation budget = HIGH:
     └─ CHOOSE: Open-minded commitment
        • Perceptual cue: Monitor for better opportunities
        • Trigger reconsideration: Goal deprioritization OR impossibility
  └─ IF Computation budget = LOW:
     └─ CHOOSE: Single-minded commitment
        • Perceptual cue: Focus on clear failure signals only

IF Environment volatility = HIGH (changes faster than typical plans):
  └─ IF Goal urgency = CRITICAL:
     └─ CHOOSE: Reactive execution (bypass BDI deliberation)
        • Perceptual cue: Environmental state requires immediate response
  └─ ELSE:
     └─ CHOOSE: Open-minded with frequent reconsideration windows
        • Perceptual cue: Schedule reconsideration at plan checkpoints

Decision Tree: Plan Library Organization Strategy

IF Domain knowledge = COMPLETE AND STABLE:
  └─ Pre-compile all plans with full context conditions
  └─ Use hash-table indexing on event types

IF Domain knowledge = INCOMPLETE OR EVOLVING:
  └─ Use hierarchical decomposition with abstract plans
  └─ Enable runtime plan composition
  └─ Index on goal patterns, not specific events

IF Real-time constraints = CRITICAL:
  └─ Sacrifice completeness for speed
  └─ Cache frequent plan instantiations
  └─ Use compiled pattern matching over full search

Decision Tree: Belief Update Granularity

IF Environmental change rate > Belief update rate:
  └─ Use event-triggered updates only
  └─ Accept temporary inconsistency for speed

IF Belief inconsistency tolerance = LOW:
  └─ Implement belief revision with dependency tracking
  └─ May require pausing execution during updates

IF Memory constraints = TIGHT:
  └─ Maintain only current-state beliefs
  └─ Compile temporal dependencies into plan preconditions

Failure Modes

1. Thrashing Agent (Continuous Reconsideration)

Symptoms: Agent switches between plans rapidly, never completing goals, high deliberation overhead Detection Rule: If reconsideration frequency > 1 per plan step execution, you're thrashing Root Cause: "Potentially significant change" detection is too sensitive OR commitment strategy is too weak for environment Fix: Tighten change detection criteria, increase commitment strength, or batch environmental updates

2. Zombie Plans (Blind Persistence)

Symptoms: Agent continues executing obviously obsolete plans, ignores contradictory evidence, goals never achieved Detection Rule: If plan execution continues after preconditions become false, you have zombie plans Root Cause: Missing "impossibility" detection in termination conditions OR belief update failures Fix: Add explicit plan precondition monitoring, implement belief-intention consistency checks

3. Option Generation Bottleneck

Symptoms: Long delays before any plan selection, agent appears "frozen" before acting, timeout failures Detection Rule: If time-to-first-action > environment change period, option generation is the bottleneck Root Cause: Plan library too large for real-time search OR matching algorithm is naive Fix: Index plans by triggering events, use compiled pattern matching, accept incompleteness for speed

4. Schema Bloat (Over-Detailed Plans)

Symptoms: Plan library grows exponentially, new situations require entirely new plans, brittle to minor variations Detection Rule: If adding new capability requires modifying >10% of existing plans, you have schema bloat Root Cause: Plans encode too much detail rather than using hierarchical decomposition Fix: Use abstract plans with subgoal decomposition, separate invariant patterns from situation-specific details

5. Desire-Intention Confusion

Symptoms: Agent attempts impossible combinations, violates resource constraints, goals conflict in execution Detection Rule: If multiple intentions require mutually exclusive resources, desires and intentions are confused Root Cause: Deliberation process doesn't filter desires for mutual consistency before commitment Fix: Implement explicit compatibility checking in deliberation, maintain resource allocation tracking

Worked Examples

Example 1: Robot Navigation Under Deadline Pressure

Scenario: Delivery robot must reach destination in 10 minutes. Environment has pedestrians (medium volatility) and network connectivity for map updates (computation budget = medium).

Decision Process:

1. Commitment Strategy Selection: Medium volatility + medium budget → Open-minded commitment
2. Initial Plan: Direct path using A* with current map
3. Execution: At waypoint 3, pedestrian blocks path
4. Reconsideration Trigger: Blocked path = "potentially significant change" because it affects plan feasibility
5. New Deliberation:
   • Option 1: Wait for pedestrian (risks deadline)
   • Option 2: Detour via loading dock (adds 2 minutes)
   • Choose: Detour (keeps deadline feasible)
6. Continued Execution: At waypoint 7, network update shows construction blocking detour
7. Reconsideration: Construction = significant change → New deliberation finds alternate route

Key Insight: Agent commits to paths but reconsiders when assumptions (clear route) become invalid. Novice would either replan at every pedestrian (thrashing) or ignore the blocked path (zombie plan).

Example 2: Trading Algorithm Under Market Volatility

Scenario: Algorithm trading in options market, goal is profit maximization, market shows high volatility.

Decision Process:

1. Environment Analysis: High volatility + profit goal (clear but moving target) → Open-minded commitment with frequent reconsideration
2. Initial Plan: Buy puts on overvalued stock XYZ
3. Execution: Places orders
4. Reconsideration Window: Every 30 seconds (predetermined based on typical option price movement)
5. Window 1: XYZ down 2%, puts profitable → Continue plan
6. Window 2: News breaks: XYZ merger announced → Stock will gap up
7. Reconsideration: News = significant change → Goal no longer achievable with current plan
8. New Deliberation: Exit put position, consider call options or different stock

Failure Trace: Without proper reconsideration windowing, agent either:

• Reacts to every price tick (thrashing, transaction costs destroy profit)
• Ignores merger news (zombie plan, massive losses)

Example 3: Satellite Control Under Communication Delays

Scenario: Earth station controlling satellite with 3-second communication delay, goal is maintain orbital position.

Decision Process:

1. Constraint Analysis: Communication delay means environment state is always 3 seconds stale, high cost of deliberation (each command cycle = 6 seconds minimum)
2. Commitment Strategy: Single-minded commitment (can't afford frequent reconsideration)
3. Plan Structure: Predictive control using orbital mechanics model
4. Execution: Send thruster commands based on predicted position
5. Reconsideration Trigger: Only when telemetry shows prediction error > safety threshold
6. Example Reconsideration: Atmospheric drag higher than predicted → orbital decay faster than expected
7. New Plan: Increase thruster frequency to compensate

Expert vs Novice:

• Novice: Tries to react to real-time telemetry → always 3 seconds behind, satellite drifts
• Expert: Uses predictive model with error-based reconsideration → maintains stable orbit despite delay

Quality Gates

• [ ] Commitment strategy explicitly selected based on environment volatility, goal clarity, and computation budget
• [ ] "Potentially significant change" detection rules defined with specific triggering conditions
• [ ] Plan library indexed for sub-second option generation in target domain
• [ ] Belief-intention consistency checks prevent impossible commitments
• [ ] Deliberation process filters desires for mutual compatibility before intention adoption
• [ ] Reconsideration frequency measured and falls within acceptable bounds (not thrashing, not zombie)
• [ ] Plan preconditions accurately reflect real-world applicability conditions
• [ ] Resource constraints (time, memory, network) explicitly modeled in architecture
• [ ] Failure modes have monitoring and recovery procedures
• [ ] Abstract interpreter semantics preserved despite implementation approximations

Not-For Boundaries

Do NOT use BDI for:

• Static optimization problems → Use mathematical programming instead
• Pure reactive control → Use behavior-based architectures instead
• Domains requiring provable optimality → Use decision theory or game theory instead
• Systems with unlimited computation time → Use classical planning instead

Delegate to other skills:

• For multi-agent coordination → Use distributed consensus protocols instead
• For learning and adaptation → Use reinforcement learning architectures instead
• For uncertainty quantification → Use probabilistic reasoning frameworks instead
• For real-time guarantees → Use real-time systems design instead