multi-agent-systems_skill

Overview

This skill helps design and implement multi-agent LLM architectures using an orchestrator-subagent pattern. It guides when to choose multi-agent vs single-agent approaches, how to isolate context, parallelize research, specialize tool sets, and build verification subagents. The focus is practical patterns, trade-offs, and concrete boundaries for effective decomposition.

How this skill works

The skill provides patterns and code sketches for orchestrator-led workflows where a lead agent decomposes tasks and delegates to specialized subagents. Subagents run isolated contexts, return concise summaries, and the orchestrator synthesizes results. It covers context-protection, parallel research, tool specialization, and verification subagents with explicit instruction templates to reduce context pollution and coordination overhead.

When to use it

• When subtasks generate large irrelevant context and only summaries are needed (context protection).
• When independent facets can be researched in parallel to explore a larger search space.
• When an agent has many tools or conflicting behavioral modes and needs specialization.
• When building a verification subagent to run tests or validate outputs.
• When clear, context-centric decomposition boundaries exist (clean API contracts, independent components).

Best practices

• Start with a single agent; add agents only when evidence shows benefits outweigh token/coordination costs.
• Decompose by context, not by generic roles or sequential phases to avoid handoff friction.
• Keep subagent outputs concise (50–200 tokens) and return structured summaries for handoffs.
• Define concrete verification criteria and require full test executions for validation agents.
• Limit tool overlap between specialists; give each agent a focused toolset and system prompt.

Example use cases

• Customer support: order lookup subagent extracts a short order summary to avoid polluting main context.
• Research project: orchestrator decomposes a topic and runs multiple research subagents concurrently, then synthesizes findings.
• Large toolset split: separate CRM and Marketing specialists each with 8–10 domain tools, routed by an orchestrator.
• Code delivery: coding agent implements feature, verification subagent runs full test suite and reports pass/fail with issues.
• Compliance workflows: specialized compliance agent enforces rigid rules while others remain flexible for brainstorming.