Multi-Agent Coordination Skill

Purpose

Provides a production-grade framework for designing and operating multi-agent systems (MAS) where two or more AI agents collaborate to decompose, execute, and synthesize complex tasks. This skill defines orchestration topologies (centralized, decentralized, hierarchical), inter-agent communication standards, DAG-based workflow decomposition, shared state architectures, compounding failure rate mitigation strategies, agent role specialization patterns, and consensus/conflict resolution protocols. The framework ensures that multi-agent deployments maintain coherence, avoid redundant work, handle partial failures gracefully, and produce synthesized outputs that exceed what any single agent could achieve alone.

Core Principles

1. Explicit Task Decomposition: Every complex task must be decomposed into a directed acyclic graph (DAG) of subtasks before delegation. Never allow agents to self-decompose without orchestrator validation.
2. Bounded Agent Autonomy: Each agent operates within a defined capability envelope. Agents must not exceed their role boundaries or invoke tools outside their authorized scope.
3. Failure Isolation: A failing agent must not cascade failures to sibling or parent agents. Circuit breakers, timeouts, and retry budgets must be enforced at every agent boundary.
4. State Convergence: All agents must converge on a single consistent view of shared state. Conflicts are resolved through deterministic merge strategies, not last-write-wins.
5. Observable Coordination: Every inter-agent message, delegation, and result must be logged with correlation IDs for end-to-end tracing across the agent graph.

Agent Protocol