agent-decision-loop-engineering

What problem does it solve? Designing an AI agent without explicit loop controls leads to runaway executions, unbounded retries, and silent low-confidence answers. This Skill provides a structured rubric for deciding how an agent's execution loop terminates, retries, and pauses for human approval, producing a reviewed design document before any code is written. ## Core Features & Use Cases - Loop Rubric Interview: Asks six batched questions covering step ceilings, termination conditions, retry policy, human-in-the-loop interrupts, ceiling-exceeded behavior, and replanning caps. - Cross-Stage Validation: Verifies that every tool marked "human approval required" in stage 4 has a corresponding interrupt point in the loop design. - Design Document Output: Writes system_design/07_loop_engineering.md with a PENDING APPROVAL status and stops for explicit user sign-off. - Use Case: While designing a LangGraph-based support agent, use this Skill to set a maximum of 15 tool calls per task, define escalation on ceiling breach, and map payment-refund tools to mandatory human approval pauses. ## Quick Start Run stage 7 of the agent-system-design pipeline to decide the agent's step ceiling, termination conditions, retry policy, and human interrupt points, then approve the generated loop engineering document.