Skip to content
pnxt

Multi-Agent Coordination Patterns

Overview

ANP envisions not a single omniscient agent but an ecosystem of specialized agents collaborating on software development. A planning agent decomposes features. Implementation agents write code. A review agent evaluates changes. A testing agent verifies behavior. A deployment agent manages releases.

This multiplicity introduces coordination challenges: How do agents divide work? How do they communicate? How do they resolve disagreements?

Topology Models

Hierarchical (Orchestrator-Worker)

        ┌───────────────┐
        │  Orchestrator  │
        └───┬───┬───┬───┘
            │   │   │
     ┌──────┘   │   └──────┐
     ▼          ▼          ▼
┌─────────┐ ┌─────────┐ ┌─────────┐
│ Worker A │ │ Worker B │ │ Worker C │
└─────────┘ └─────────┘ └─────────┘

Best for: Well-defined tasks that decompose into independent subtasks. Clear authority, simple coordination, but single point of failure.

Peer-to-Peer (Collaborative)

┌─────────┐     ┌─────────┐
│ Agent A  │◄───▶│ Agent B  │
└────┬─────┘     └────┬─────┘
     │    ┌─────────┐ │
     └───▶│ Agent C  │◀┘
          └─────────┘

Best for: Creative and exploratory tasks. No single point of failure, but coordination complexity grows quadratically.

Hybrid

Combines hierarchical structure for task decomposition with peer communication for collaboration within subteams. This is the recommended topology for most real-world scenarios.

Task Decomposition

The document catalogs strategies for breaking work into agent-assignable units:

  • Functional decomposition — By module or feature boundary
  • Temporal decomposition — By workflow phase (plan → implement → review → test)
  • Skill-based decomposition — By agent specialization
  • Dependency-aware decomposition — Minimizing cross-agent dependencies

Communication Patterns

  • Request-Response — Direct task delegation with results
  • Publish-Subscribe — Broadcasting events (e.g., “file changed”, “tests passed”)
  • Shared Blackboard — Common workspace for collaborative problem-solving
  • Negotiation — Agents propose, counter-propose, and agree on approach

Conflict Resolution

When agents disagree:

  1. Structural prevention — Clear boundaries prevent most conflicts
  2. Automated resolution — Merge strategies, version control
  3. Escalation — Unresolvable conflicts escalate to orchestrator or human

Key Contribution

Match topology to task type, delegate completely to minimize coordination overhead, and prevent conflicts through clear boundaries rather than resolving them after the fact.