multi-agent-patterns_skill

Overview

This skill explains how to design and implement multi-agent architectures focused on context isolation, reliable coordination, and production readiness. It presents supervisor, peer-to-peer (swarm), and hierarchical patterns, with practical trade-offs and mitigations for common failure modes. Use it to scale reasoning beyond single-context limits and to structure agents for parallel work and specialization.

How this skill works

The skill inspects task decomposition, context partitioning, and coordination protocols to recommend architectural patterns and implementation tactics. It evaluates when to isolate context, when to delegate full context vs. instruction passing, and how to route results through supervisors, direct pass-throughs, or peer handoffs. It also prescribes consensus methods (weighted voting, debate) and operational safeguards like TTLs, checkpointing, and validation gates.

When to use it

• When single-agent context limits constrain task complexity or accuracy
• When tasks decompose into parallel subtasks requiring separate prompts or tools
• When different subtasks need specialized toolsets or system instructions
• When building production systems that must scale across domains or users
• When you need robust coordination, convergence checks, or failure isolation

Best practices

• Design for context isolation first: sub-agents should minimize carried state
• Choose pattern by coordination need, not organizational metaphor (supervisor, swarm, hierarchical)
• Implement explicit handoff protocols and state-passing formats to avoid ambiguity
• Use weighted voting or debate protocols to avoid consensus on hallucinations
• Enforce output schemas, validation, and retry/circuit-breakers before passing results
• Batch and async communication where possible to reduce token/latency overhead

Example use cases

• Research pipeline: supervisor routes search, analysis, fact-check, and writing agents with direct pass-throughs for finalized sections
• Customer support: handoff protocol routes issues to billing, technical, or sales agents with preserved request metadata
• Parallel document analysis: multiple analyzer agents process documents in parallel and a planner aggregates results with convergence checks
• Enterprise workflow: hierarchical layers (strategy → planning → execution) manage long-running multi-step projects with checkpointing