What problem does it solve?
Uncoordinated AI agent engineering workflows often lead to inconsistent output quality, uncontrolled model costs, and undetected regression risks from unvalidated implementation changes. This Skill provides a structured framework to align agent execution with human quality and risk controls.
Core Features & Use Cases
- Eval-First Execution Loop: Define capability and regression evals upfront to catch performance drops and unintended changes immediately after implementation.
- Task Decomposition & Model Routing: Break work into 15-minute independently verifiable units, and route tasks to the optimal model tier (Haiku for simple edits, Sonnet for implementation, Opus for architecture) to balance cost and performance.
- Structured Review & Cost Discipline: Focus human review on high-risk areas like invariants, edge cases, and security assumptions, while tracking per-task model usage, token costs, and success rates to optimize spending.
- Use Case: When building a new feature with AI agents, use this Skill to define evals for the feature, decompose work into small testable units, avoid wasting review time on style issues, and only escalate to more expensive models when lower tiers fail with clear reasoning gaps.
Quick Start
Use the agentic-engineering skill to plan and execute the implementation of a new user authentication feature with built-in eval checks and cost-aware model routing.
