| name | agentic-engineering |
| description | Use when building or debugging LangGraph multi-agent systems - eval-first execution, task decomposition, model routing by complexity, and cost discipline |
Agentic Engineering
For engineering workflows where AI agents perform most implementation work and humans enforce quality.
Operating Principles
- Define completion criteria before execution
- Decompose work into agent-sized units (15-min rule)
- Route model tiers by task complexity
- Measure with evals and regression checks
Model Routing
| Task Complexity | Model Tier | Examples |
|---|
| Classification, boilerplate, narrow edits | Fast/cheap | Tag routing, format conversion |
| Implementation and refactors | Standard | API endpoints, DB queries, tests |
| Architecture, root-cause, multi-file invariants | Powerful | ORM design, security review, debugging |
Task Decomposition (15-min unit rule)
- Each unit: independently verifiable, single dominant risk, clear done condition
- Closely-coupled units: same session
- Phase transitions: start fresh session
Review Focus for AI-Generated Code
Prioritize (don't waste cycles on style):
- Invariants and edge cases
- Error boundaries and rollback
- Security and auth assumptions
- Hidden coupling and rollout risk
Cost Discipline
Track per task: model, token estimate, retries, wall-clock, success/failure.
Escalate model tier only when lower tier fails with clear reasoning gap.
Session Strategy
- Compact after milestone completion, not during active debugging
- Continue session for closely-coupled units
- Start fresh after major phase transitions
