curiositech/resource-bounded-planning

SKILL: Resource-Bounded Planning and Reasoning

Core Purpose

Design and implement agents that balance deliberation depth against execution urgency — making "good enough" decisions under time pressure while maintaining adaptability to change.

DECISION POINTS

1. Commitment Formation Threshold

IF: Domain has recurring decisions + Time pressure exists + Information is incomplete
├─ High uncertainty about environment changes
│  └─ Commit to high-level structure only (defer implementation details)
├─ Moderate uncertainty with recurring patterns  
│  └─ Commit to partial plan with 2-3 refinement checkpoints
└─ Low uncertainty with established patterns
   └─ Commit to detailed plan with monitoring triggers

Cost-Benefit Formula: Commit when (Expected_Deliberation_Savings × Recurrence_Rate) > (Revision_Cost × Uncertainty_Factor)

2. Filter Override Calibration

IF: New incompatible option detected
├─ Expected value > Current plan value + (Override_threshold × Deliberation_cost)
│  └─ TRIGGER: Begin costly reconsideration
├─ Expected value > Current plan value but < Override threshold
│  └─ FILTER: Log opportunity but maintain current commitment  
└─ Expected value ≤ Current plan value
   └─ FILTER: Ignore option completely

Target Override Frequency: 10-20% of detected options (adjust threshold accordingly)

3. Structural Partiality Levels

IF: Planning horizon analysis needed
├─ Execution time < 24 hours
│  └─ Specify to action level (full detail required)
├─ Execution time 1-30 days  
│  └─ Specify to task level (keep method flexibility)
├─ Execution time > 30 days
│  └─ Specify to goal level only (maximum adaptability)

Refinement Trigger: When means-end coherence gap blocks next action OR dependency requires specification

4. Multi-Agent Consistency Checking

IF: Agent coordination required
├─ Shared resource conflicts detected
│  └─ Negotiate commitment revision (cost = deliberation + coordination overhead)
├─ Goal conflicts detected but no resource overlap
│  └─ Maintain separate commitments with monitoring
└─ Complementary goals detected
   └─ Establish consistency constraints (shared assumptions + checkpoint synchronization)

Consistency Check Frequency: Every N actions where N = 1/(Coordination_criticality × Change_rate)

FAILURE MODES

1. Perpetual Deliberation (Situation 2b/3)

Detection Rule: If deliberation time > 2× action time OR override frequency > 30% of options Symptoms: Constantly reconsidering plans, thrashing between options, missing deadlines Diagnosis: Override threshold too low, filter mechanism too weak Fix: Increase override threshold by 20-30%, strengthen filter criteria, set hard deliberation deadlines

2. Brittle Automation (Situation 4a)

Detection Rule: If plan execution fails due to assumption violations > 2× per planning cycle Symptoms: Missing significant opportunities, executing invalid plans, inflexible to change Diagnosis: Override threshold too high, monitoring inadequate Fix: Decrease override threshold by 15-25%, add assumption monitoring, implement adaptation triggers

3. Premature Specification

Detection Rule: If plan details change > 40% before execution OR refinement cost > initial planning cost Symptoms: Wasted computation on obsolete details, reduced flexibility, high revision overhead Diagnosis: Committing to unnecessary detail too early Fix: Defer specification until means-end coherence requires OR execution approaches

4. Consistency Cascade Failures

Detection Rule: If coordination overhead > 25% of productive work time Symptoms: Agents constantly negotiating, inconsistent assumptions across agents, deadlock states Diagnosis: Consistency requirements too strict OR communication protocols inefficient Fix: Relax non-critical consistency constraints, implement local consistency checking, batch coordination updates

5. Analysis Paralysis Spiral

Detection Rule: If deliberation cost exceeds action value for same decision type > 3 consecutive times Symptoms: Over-analyzing low-stakes decisions, missing action windows, computational resource waste Diagnosis: Treating all decisions as equally important Fix: Implement decision importance classification, set deliberation budgets per decision class, use satisficing for low-stakes choices

WORKED EXAMPLES

Example 1: Multi-Agent Resource Allocation Under Time Pressure

Scenario: Three AI agents managing a data center during peak load. Agent A (scheduler), Agent B (resource allocator), Agent C (performance monitor). Sudden 300% traffic spike detected.

Novice Approach: Each agent recomputes optimal strategy independently

• Result: 45 seconds deliberation, conflicting resource assignments, system degraded

Expert Application:

1. Commitment Formation (T+2 seconds):

   • Agent A: Commits to "shed 40% low-priority traffic" (high-level only)
   • Agent B: Commits to "reallocate 60% compute to web tier" (defers specific VM assignments)
   • Agent C: Commits to "monitor latency < 200ms threshold" (specific metric, flexible response)

2. Consistency Check (T+3 seconds):

   • Shared assumption: Traffic spike temporary (< 30 min)
   • Constraint: No agent action should block others' execution
   • Checkpoint: Reassess at T+10 minutes

3. Execution with Refinement (T+4 to T+600):

   • Agent A starts shedding traffic immediately (acts on partial plan)
   • Agent B refines allocation as A's actions free up capacity
   • Agent C detects threshold breach at T+180, triggers coordinated adjustment

4. Override Decision (T+300):

   • New option detected: Emergency capacity from partner DC (high value but high deliberation cost)
   • Override threshold analysis: Expected value (300 units) > Current plan (200) + (50 threshold × 40 deliberation cost) = 2200
   • Decision: Filter the option (below threshold), maintain current commitments

Measurable Outcomes:

• System stabilized at T+240 seconds (vs. T+420 with recomputation)
• Override triggered only once (8% of detected options)
• Coordination overhead: 12% of total response time

Example 2: Hierarchical Task Decomposition in Software Development

Scenario: AI coding assistant planning implementation of new feature. Requirements partially specified, 2-week deadline, dependencies uncertain.

Trade-off Analysis:

Option 1: Fully specify all implementation details upfront

• Pro: Clear execution path
• Con: 80% of details likely to change, high revision cost
• Expected cost: 40 hours planning + 60 hours revision = 100 hours

Option 2: Commit only to architecture, defer all implementation

• Pro: Maximum flexibility
• Con: Cannot parallelize work, means-end coherence gaps block progress
• Expected cost: 10 hours planning + 120 hours sequential implementation = 130 hours

Expert Choice - Option 3: Structural partiality with refinement triggers

• Commit immediately: API interface design, data model schema
• Defer: Implementation algorithms, UI details, error handling specifics
• Refinement triggers: When frontend team needs UI specs (Week 1), when testing reveals performance issues

Execution:

1. T+0: Form partial plan (4 hours)

   • Commit: RESTful API with 6 endpoints, PostgreSQL schema
   • Defer: Caching strategy, authentication details, frontend components

2. T+3 days: First refinement trigger

   • Means-end coherence gap: Frontend blocked on authentication flow
   • Refine: OAuth integration details (2 hours)
   • Still defer: Caching, error handling

3. T+8 days: Performance monitoring trigger

   • Assumption violation: API response time > 500ms
   • Override decision: Deliberation cost (8 hours) vs. Performance value (high) → Trigger override
   • Add: Redis caching layer, refactor 3 endpoints

Quality Results:

• Total planning time: 14 hours (vs. 100 or 130)
• Requirements changes handled: 7 (with minimal rework due to deferred details)
• Delivery: On schedule with 95% requirements satisfaction

QUALITY GATES

Planning Quality Checklist:

• [ ] Commitment level explicitly specified (high-level/task-level/action-level)
• [ ] Filter override threshold quantified with cost-benefit formula
• [ ] Structural partiality justified (what deferred and why)
• [ ] Refinement triggers identified with objective criteria
• [ ] Assumption monitoring mechanisms in place
• [ ] Deliberation budget set and tracked
• [ ] Override frequency target established (10-20% baseline)
• [ ] Consistency constraints defined for multi-agent scenarios
• [ ] Failure mode detection rules implemented
• [ ] Plan revision cost estimated and acceptable

Execution Quality Checklist:

• [ ] Acting on partial plans while continuing refinement
• [ ] Override decisions documented with threshold calculations
• [ ] Deliberation interruption occurring at deadlines
• [ ] Means-end coherence gaps triggering appropriate refinement
• [ ] Consistency violations detected within 1 cycle
• [ ] Filter mechanism blocking > 70% of incompatible options
• [ ] Revision frequency within expected range for uncertainty level

Calibration Quality Indicators:

• [ ] Override frequency 10-20% (if >30%: threshold too low; if <5%: too high)
• [ ] Deliberation cost < 25% of total task time
• [ ] Plan revision frequency matches environmental change rate ±20%
• [ ] Coordination overhead < 20% of productive work time
• [ ] Assumption violations caught before execution failure

NOT-FOR BOUNDARIES

This skill is NOT for:

• Complete information scenarios: Use classical decision theory instead
• Single-shot decisions with unlimited time: Use optimization algorithms instead
• Deterministic environments with fixed requirements: Use traditional planning instead
• Systems where deliberation cost is negligible: Use search-based approaches instead
• Perfect coordination requirements: Use centralized control instead

Delegate to other skills when:

• Probabilistic reasoning needed: Use [bayesian-reasoning] for uncertainty quantification
• Game-theoretic scenarios: Use [strategic-interaction] for multi-agent competition
• Real-time control systems: Use [control-theory] for continuous feedback loops
• Knowledge representation: Use [knowledge-engineering] for complex domain modeling
• Learning from experience: Use [reinforcement-learning] for policy improvement over time

Warning signs you're misapplying this skill:

• Spending more time on meta-reasoning about planning than on planning itself
• Treating all decisions as equally resource-constrained
• Applying override mechanisms to deterministic rule-following
• Using structural partiality when complete specification is both possible and stable
• Implementing BDI architecture for single-threaded, batch processing tasks