What problem does it solve? LLM agents fail in predictable ways — infinite loops, tool misuse, runaway costs, and prompt-injection-driven actions. This Skill provides a structured design and review playbook that addresses these failure modes up front, helping you decide whether a task needs an agent at all and how to constrain it when it does. ## Core Features & Use Cases - Autonomy Escalation Ladder: Decide between a single LLM call, a fixed chain, an agent, or multi-agent orchestration before writing any code. - Tool & Loop Design Rules: Guidelines for orthogonal tool sets, description-as-prompt authoring, hard iteration/token/time caps, loop detection, and explicit terminal states. - Guardrails & Evaluation Checklists: Injection canary tests, code-level gating of destructive actions, per-run tracing, and eval suites with task fixtures. - Use Case: When your support agent keeps retrying the same failing tool call and burning API budget, use this Skill's loop-containment and review checklist to add caps, loop detection, and explicit failure states. ## Quick Start Ask the AI to review your existing agent implementation against the agent-design checklist and identify missing loop caps, ungated destructive tools, and injection vulnerabilities.
