PUMA Vault

An Introduction to MultiAgent Systems

Se lee en 3 min

LN: Wooldridge (2009) — An Introduction to MultiAgent Systems (2nd ed.)

Bibliographic Reference

Citation: Wooldridge, M. (2009). An introduction to multiagent systems (2nd ed.). John Wiley & Sons. https://www.wiley.com/en-us/An+Introduction+to+MultiAgent+Systems%2C+2nd+Edition-p-9780470519462

Pass 1 — Bird’s Eye View (5 Cs)

CAssessment
CategoryAcademic textbook — foundational reference
ContextMichael Wooldridge (Oxford, formerly Liverpool) is one of the founding figures of MAS research. This textbook is the standard academic reference for the field
CorrectnessPeer-reviewed; covers formal foundations (BDI logic, game theory) alongside practical architectures
Contributions(1) Formal definition of agents and MAS; (2) BDI (Belief-Desire-Intention) agent architecture; (3) Reactive vs. deliberative agents; (4) Coordination, cooperation, and negotiation protocols; (5) Agent communication languages; (6) Multi-agent planning
ClarityExcellent — balances formal rigour with accessible explanations

Relevance: ⭐⭐⭐⭐⭐

The standard academic definition of agents and MAS that underpins PUMA’s agent architecture. The Manager Agent, Triage Agent, and QA Agent in BMAD are MAS agents in Wooldridge’s sense. Justifies PUMA’s multi-agent design as academically grounded.

Pass 2 — Key Concepts

Formal Agent Definition

“An agent is a computer system that is situated in some environment and that is capable of autonomous action in this environment in order to meet its delegated objectives.”

Key properties of agents:

  • Autonomy: Operates without direct intervention; has control over its own state
  • Social ability: Interacts with other agents via an agent communication language
  • Reactivity: Perceives and responds to its environment
  • Pro-activeness: Takes initiative; pursues goals, not just responses to stimuli

Reactive vs. Deliberative Agents

TypeArchitectureStrengthWeakness
ReactiveStimulus-response rulesFast, no overheadNo planning, no goals
Deliberative (BDI)Belief-Desire-IntentionGoal-directed, plansComputational overhead
HybridReactive layer + deliberative layerBest of bothDesign complexity

PUMA’s agents are BDI-inspired: they have beliefs (context from the issue + prompt), desires (correct classification), and intentions (the chain-of-thought leading to the output).

BDI Architecture

Beliefs: What the agent knows about the world
         (issue description, priority schema, few-shot examples)
Desires: What the agent wants to achieve
         (correct triage classification, accurate SP estimate)
Intentions: What the agent is committed to doing
            (current CoT reasoning chain → output)

Coordination and Negotiation

In MAS with multiple agents:

  • Coordination: Ensuring agents don’t conflict or duplicate work
  • Cooperation: Agents work together towards shared goals
  • Negotiation: Agents reach agreements on resource allocation

In PUMA’s BMAD multi-agent team:

  • Manager Agent coordinates task allocation
  • Analyst Agent + Triage Agent cooperate on issue classification
  • QA Agent independently validates outputs (adversarial coordination)

Agent Communication

Wooldridge formalises agent communication using KQML (Knowledge Query and Manipulation Language) and FIPA ACL. Modern LLM-agent frameworks (LangGraph, CrewAI) implement these principles through structured message passing and shared state.

PUMA Integration

Wooldridge provides the academic foundation for PUMA’s multi-agent design:

Wooldridge ConceptPUMA Implementation
Agent autonomyEach BMAD agent operates independently with defined scope
Reactive layerIssue ingestion pipeline (immediate classification)
Deliberative layerCoT reasoning for complex issues
CoordinationManager Agent as coordinator
CooperationTriage + Analyst cooperative classification
NegotiationHITL escalation as human-agent negotiation

MOCs

Vista Gráfica

Retroenlaces