curiositech/windags-skills

Wooldridge Multi-Agent Systems

Use this skill when the question is not "how do I make several models talk," but "what kind of autonomous system does this environment force me to build?"

When to Use

• You must choose between reactive, deliberative, or hybrid agent architectures.
• Coordination keeps failing because no one has complete information or uncontested control.
• Agents need negotiation, allocation, or commitment strategies under uncertainty.
• A design depends on what agents know, what everyone knows, or whether common knowledge is even achievable.
• You need to diagnose whether a "multi-agent" design is actually just asynchronous object orchestration.

NOT for

• Single-agent planning or optimization problems with no coordination requirement.
• Centralized schedulers where one controller already owns global decision rights.
• Prompt-only roleplay swarms that do not have separate goals, beliefs, or control boundaries.
• General distributed-systems design when agent autonomy and knowledge asymmetry are irrelevant.

Core Mental Models

Environment Properties Drive Architecture

Start with observability, determinism, dynamics, and time pressure. Architecture choice is downstream of environment shape, not personal preference for a fashionable agent pattern.

Autonomy Is Control Inversion

Agents are not just asynchronous objects. They decide whether and when to comply, which means requests need semantics, refusals, and coordination logic rather than implicit obedience.

Coordination Emerges from Constraint

Partial observability, resource contention, and interdependent goals create coordination pressure. Good protocol design begins by naming that pressure instead of assuming collaboration is always desirable.

Commitment Strategy Must Match Volatility

Bold agents overcommit in fast-changing environments; cautious agents thrash in stable ones. Reconsideration frequency is a design parameter tied to environment dynamics, not a universal best practice.

Knowledge Levels Matter

Individual knowledge, everyone-knows, common knowledge, and distributed knowledge are not interchangeable. Asking for the wrong level can make a protocol impossible or wastefully expensive.

Decision Points

flowchart TD
  A[Need agent architecture] --> B{Environment fully observable and stable?}
  B -->|Yes| C[Simple reactive or planner-heavy design may suffice]
  B -->|No| D{Need real-time reaction under uncertainty?}
  D -->|Yes| E[Hybrid architecture]
  D -->|No| F[Deliberative architecture]
  E --> G{Why do agents need to coordinate?}
  F --> G
  G -->|Partial information| H[Information-sharing protocol]
  G -->|Resource contention| I[Negotiation or allocation mechanism]
  G -->|Task dependency| J[Commitment or delegation protocol]
  G -->|No real autonomy| K[Use a simpler centralized design]

• Characterize the environment before choosing an agent architecture.
• If agents cannot refuse, delay, or reinterpret requests, do not pretend you have agent autonomy.
• Choose knowledge requirements intentionally. Many systems need distributed knowledge, not common knowledge.
• Tune commitment boldness to the environment's change rate and the cost of reconsideration.

Failure Modes

Architecture-First Design

Cue: the team wants BDI, debate, or a fancy agent framework before anyone can describe the environment.

Fix: force environment characterization first.

Anthropomorphic Autonomy

Cue: the design talks about beliefs and desires but cannot map them to local state, observations, or protocol obligations.

Fix: ground mental language in concrete computational state.

Coordination for Its Own Sake

Cue: communication traffic rises, but no one can explain which environmental pressure requires it.

Fix: identify the actual forcing function and design the minimal protocol that addresses it.

Hidden Central Controller

Cue: one "agent" silently makes all important choices while the others just execute.

Fix: either admit the architecture is centralized or redistribute genuine decision rights.

Impossible Knowledge Assumptions

Cue: protocol correctness depends on every agent knowing that every other agent knows a fact, but the network is unreliable.

Fix: relax to distributed knowledge, acknowledgments, or eventual consistency as the environment allows.

Worked Examples

Specialist LLM Review Swarm

A coding system uses security, performance, and style agents. Security reviews depend on different evidence than style review, and no single reviewer has global visibility. Use a hybrid coordinator plus targeted information-sharing instead of assuming every reviewer should see the entire context.

Distributed Vehicle Monitoring

Sensor agents cover overlapping but incomplete regions. Coordination is necessary because no single agent can maintain end-to-end track continuity. Design the protocol around partial observability rather than generic "collaboration."

Quality Gates

• The environment is characterized explicitly before architecture choice.
• The design names the real source of coordination pressure.
• Knowledge requirements are stated at the right level.
• Commitment strategy is tied to volatility and reconsideration cost.
• Any claimed autonomy includes refusal, delay, or local interpretation of requests.

Shibboleths

• If someone calls a system multi-agent but every important decision still routes through one controller, they are renaming distributed execution, not designing autonomy.
• If "common knowledge" is used casually with no communication model, the protocol is probably underspecified.
• If the team cannot say why agents must coordinate, the design probably does not need agents at all.

Reference Routing

• references/environment-characterization-drives-architecture.md: load when architecture choice is the main question.
• references/coordination-as-necessity-not-luxury.md: load when you must justify or minimize coordination.
• references/commitment-strategies-and-environment-dynamics.md: load when boldness or replanning cadence is the main tuning issue.
• references/grounded-epistemic-logic-for-distributed-agents.md: load when knowledge claims drive safety or correctness.
• references/negotiation-and-resource-allocation-mechanisms.md: load when autonomy collides with scarce resources.