What problem does it solve? Teams frequently propose multi-agent designs (subagents, swarms, orchestrators, agent handoffs) without evidence that the added coordination cost is justified, leading to higher token spend, compounding errors, and debugging difficulty. This Skill challenges those proposals with a structured five-gate test and published cost and failure data. ## Core Features & Use Cases - Five-Gate Evaluation: Checks whether a proposed multi-agent design has genuinely independent threads, read-only or cleanly partitioned writes, evidence exceeding one context window, value absorbing a ~15x token multiplier, and a verification pass on synthesis. - Cost and Failure Evidence: Cites Anthropic's token-multiplier data, Berkeley's MAST study (41%-86.7% failure rates across seven frameworks), and the read/write asymmetry rule to argue against unnecessary fan-out. - Topology Ranking: Rates orchestrator-worker with read-only workers, sequential pipelines, hierarchical, and peer-network topologies, and identifies portable single-agent patterns like clean-context review. - Use Case: When a design document proposes splitting a single agent into an orchestrator with six workers, run this review to test the split against the five gates and present the single-agent-with-larger-budget control comparison. ## Quick Start Review this proposed multi-agent architecture and argue against it using the five-gate test unless every gate passes.
