Agentic Engineering

1---2name: agentic-engineering3description: Operate as an agentic engineer using eval-first execution, decomposition, and cost-aware model routing.4origin: ECC5---6 7# Agentic Engineering8 9Use this skill for engineering workflows where AI agents perform most implementation work and humans enforce quality and risk controls.10 11## Operating Principles12 131. Define completion criteria before execution.142. Decompose work into agent-sized units.153. Route model tiers by task complexity.164. Measure with evals and regression checks.17 18## Eval-First Loop19 201. Define capability eval and regression eval.212. Run baseline and capture failure signatures.223. Execute implementation.234. Re-run evals and compare deltas.24 25## Task Decomposition26 27Apply the 15-minute unit rule:28- each unit should be independently verifiable29- each unit should have a single dominant risk30- each unit should expose a clear done condition31 32## Model Routing33 34- Haiku: classification, boilerplate transforms, narrow edits35- Sonnet: implementation and refactors36- Opus: architecture, root-cause analysis, multi-file invariants37 38## Session Strategy39 40- Continue session for closely-coupled units.41- Start fresh session after major phase transitions.42- Compact after milestone completion, not during active debugging.43 44## Review Focus for AI-Generated Code45 46Prioritize:47- invariants and edge cases48- error boundaries49- security and auth assumptions50- hidden coupling and rollout risk51 52Do not waste review cycles on style-only disagreements when automated format/lint already enforce style.53 54## Cost Discipline55 56Track per task:57- model58- token estimate59- retries60- wall-clock time61- success/failure62 63Escalate model tier only when lower tier fails with a clear reasoning gap.