ai-agent-multi-agent-coordination

# AI Multi-Agent Coordination Acknowledgement: Shared by Peter Bamuhigire, techguypeter.com, +256 784 464178. <!-- dual-compat-start --> ## Use When - A single agent cannot reliably do the job (too many tools, too many domains, ambiguous routing). - Standing up a **supervisor + workers** pattern (one planner, several specialists). - Implementing a **debate** pattern where two agents propose answers and a judge picks (or a critic refines). - Designing **plan-execute** (one plan-only agent, one execute-only agent) for separation of concerns. - Implementing **agent-to-agent handoff** (sales agent hands to billing agent mid-task). ## Do Not Use When - A single agent can do the task — use it. Multi-agent **multiplies cost and latency** and **divides reliability**. Default to single-agent. - The task is just multi-tool — that's not multi-agent; one agent with N tools is still one agent. - The task is a deterministic workflow with LLM steps — `ai-agent-runtime-architecture` §1 decision. ## Required Inputs - Single-agent design first. Only escalate to multi-agent when single-agent demonstrably fails on eval. - Agent runtime (`ai-agent-runtime-architecture`). - Tool registry (`ai-agent-tool-catalogue-and-action-gating`). - Cost / step budgets (`ai-agent-cost-and-step-budgets`) — multi-agent budgets are an order of magnitude larger. ## Workflow 1. Read this `SKILL.md`. 2. Justify multi-agent (§1). Reject it as default. 3. Choose a **coordination pattern** (§2): supervisor/worke...