What problem does it solve? Teams building LLM agents routinely over-engineer — reaching for multi-agent swarms when a single call would do — or ship agents missing basic guardrails like step budgets, tracing, and human approval gates. This Skill guides the design of a new agentic system or audits an existing one against a catalog of known failure modes. ## Core Features & Use Cases - Greenfield layered design interview: Walks an autonomy gate (single call → workflow → agent → multi-agent) then composes reasoning loop, topology, memory, reliability, governance, and integration layers, each with its named cost. - Nine-point agent code review: Audits existing agent systems for seven over-building defects, one under-building check, and one stale-scaffold check, then recommends the highest-leverage simplification. - Decision persistence: Records every recommendation — including explicit refusals like "no agent needed" — as a proposed decision file with an auditable constitution. - Use Case: Ask whether your support bot needs a multi-agent setup; the Skill maps your current design, flags ungated irreversible actions and missing step budgets, and recommends collapsing to one well-tooled agent with HITL gates. ## Quick Start Ask the agent to review my LangGraph multi-agent setup and tell me whether a single agent with tools would be sufficient.
