curiositech/smith-1980-contract-net-protocol

SKILL.md — Big Brother Logic: Epistemic Reasoning for Multi-Agent Systems

DECISION POINTS

Primary Decision Tree: Responding to Partial Information States

Given: Agent state S, Goal G, Available actions A

IF satisfiability_check(S, G) == TRUE
├── IF current_knowledge_sufficient(S, G)
│   └── Execute planned action
├── ELSE IF gather_info_viable(S, A)
│   ├── Identify minimum vision set expansion needed
│   ├── Request targeted information from other agents
│   └── Update Kripke model with new information
│   └── Re-evaluate satisfiability_check(S, G)
└── ELSE
    └── Reconfigure agent positions/permissions to satisfy G

ELSE IF satisfiability_check(S, G) == FALSE
├── IF goal_relaxation_acceptable()
│   └── Modify G to achievable subset
└── ELSE
    └── Abort and escalate to human operator

ELSE IF satisfiability_check(S, G) == UNKNOWN
├── Attempt distributed reasoning with peer agents
├── IF still UNKNOWN after coordination
│   └── Escalate to centralized knowledge computation
└── ELSE proceed with satisfiability branch above

Coordination Failure Recovery Decision Tree

When coordination attempt fails between agents A and B:

IF agents had same information but different actions
├── Check for common knowledge gap
│   ├── Was information publicly announced? → Use public broadcast
│   └── Was announcement verified received? → Add confirmation protocol
└── Check for conflicting vision sets → Resolve perceptual boundaries

IF agents had different information
├── Map each agent's vision set
├── Identify which agent has authoritative view
├── Update uninformed agent's Kripke model
└── Re-attempt coordination

IF coordination succeeds but action fails
├── Epistemic goal was wrong, not coordination
└── Revise knowledge requirements for this task type

FAILURE MODES

1. "False Knowledge" Anti-Pattern

Symptom: Agent claims to "know" something based on incomplete information Detection Rule: If agent acts on belief X but cannot rule out scenarios where ¬X, it has false knowledge Fix: Map agent's vision set; expand perceptual boundaries or add verification step before action

2. "Mutual Knowledge Masquerade"

Symptom: System assumes coordination after broadcasting message to all agents Detection Rule: If coordination fails despite "shared" information, check if agents know that others received the message Fix: Replace broadcast with public announcement protocol; verify common knowledge establishment

3. "Epistemic Goal Drift"

Symptom: Agents follow procedures correctly but system fails to achieve intended outcome Detection Rule: If tasks complete successfully but higher-level goal fails, procedures were specified without epistemic foundation Fix: Rewrite specifications as "Agent X must know Y before doing Z" instead of procedural steps

4. "Centralization Denial"

Symptom: System uses central knowledge computation but claims to be "distributed" Detection Rule: If any single point of failure can corrupt all agents' knowledge states Fix: Either accept centralized architecture with honest trade-off documentation, or redesign for true distributed epistemic reasoning

5. "Vision Set Mismatch"

Symptom: Agent assigned task requiring information outside its perceptual boundaries Detection Rule: If agent cannot distinguish scenarios relevant to its assigned task Fix: Either expand agent's vision set or reassign task to agent with appropriate perceptual access

WORKED EXAMPLES

Example 1: Agent Misconfiguration Recovery

Scenario: Three surveillance agents (A1, A2, A3) monitoring area. A2's camera malfunctions, creating coverage gap.

Initial State: A1 knows east sector clear, A3 knows west sector clear, A2 reports nothing (due to malfunction) Goal: Verify entire area is secure before allowing human entry

Decision Process:

1. satisfiability_check(current_state, "area_secure") → UNKNOWN (A2's sector unverified)
2. gather_info_viable() → FALSE (A2 cannot provide info)
3. Reconfigure: A1 and A3 adjust positions to overlap A2's sector
4. New vision sets: A1 covers east + center, A3 covers west + center
5. satisfiability_check(new_state, "area_secure") → TRUE
6. Execute: Allow human entry

Expert Insight: Novice would wait for A2 to recover or manually check the area. Expert recognizes this as satisfiability problem and solves via reconfiguration.

Example 2: Common Knowledge Coordination

Scenario: Financial trading agents must execute synchronized trades across markets

Initial Attempt: Central system broadcasts "execute trades at 14:30" to all agents Failure: Some agents execute, others don't, causing market position mismatch

Epistemic Analysis:

• Each agent received message (mutual knowledge)
• But agents don't know others received it (no common knowledge)
• Without common knowledge, coordination fails in adversarial environment

Solution:

1. Replace broadcast with public announcement requiring confirmation
2. Each agent confirms receipt and sees others' confirmations
3. Only proceed when common knowledge of "all agents ready" is established
4. Result: Perfect synchronization achieved

QUALITY GATES

Task completion requires ALL conditions satisfied:

• [ ] Vision set adequacy: Each agent can distinguish all scenarios relevant to its assigned tasks
• [ ] Common knowledge depth: For coordination-critical information, all agents know that all agents know (verified to required depth)
• [ ] Satisfiability verification: Current epistemic state provably satisfies all knowledge goals before action
• [ ] State monotonicity: New information only expands agent knowledge, never contradicts existing knowledge
• [ ] Failure mode coverage: System behavior defined for all epistemic failure cases (unknown, false knowledge, coordination failure)
• [ ] Centralization trade-off documented: If using centralized computation, failure modes and limitations explicitly stated
• [ ] Communication epistemic effect: Each message's impact on agent Kripke models formally specified
• [ ] Runtime verification active: Epistemic properties checked before irreversible actions
• [ ] Goal achievability confirmed: All epistemic goals are satisfiable given current system constraints
• [ ] Boundary condition handling: System behavior defined when goals become unsatisfiable

NOT-FOR BOUNDARIES

This skill should NOT be used for:

• Single-agent reasoning tasks → Use standard planning/decision-making frameworks
• Task coordination with complete shared information → Use workflow management systems
• Performance optimization of existing working systems → Use profiling/optimization tools
• Simple message passing between components → Use standard communication patterns
• Systems where "good enough" coordination is acceptable → Use eventual consistency patterns

Delegate to other skills when:

• Need real-time performance optimization → [performance-optimization-skill]
• Designing human-AI interaction workflows → [human-ai-collaboration-skill]
• Building fault-tolerant distributed systems → [distributed-systems-resilience-skill]
• Implementing security/access control → [multi-agent-security-skill]

Use this skill specifically when:

• Knowledge asymmetry between agents is the core challenge
• Coordination failures occur despite agents having "correct" information
• System must guarantee epistemic properties, not just attempt coordination
• Need formal verification that agents "know enough" before acting