graph-engineeringlisted

# Graph Engineering — Agent Workflow Design > **Invoke this skill whenever you are designing any multi-step LLM or agent task.** > The goal: replace fragile linear chains with resilient graphs where agents can reflect, branch, and self-correct. --- ## The Problem This Fixes Most people who build a multi-step agent end up with a **straight line**: ``` prompt → step 1 → step 2 → step 3 → output ``` When step 1 emits imperfect output, everything downstream compounds the error — silently. **Graph Engineering** turns that chain into a *network* where: - Nodes can loop and self-correct (**reflection**) - Agents can reach outside themselves for fresh data (**tool use**) - Intent is made explicit before execution (**planning**) - Responsibilities are split across specialists (**multi-agent**) - Output quality is verified against a rubric before shipping (**critique**) > Core insight from Andrew Ng's 4-agentic-patterns paper + codila's framing: > **Loops let agents think. Graphs let agents remember.** --- ## The 5 Layers Apply these in order. **Do NOT add all 5 by default** — each layer adds latency and cost. Add only what a real failure mode demands. --- ### Layer 1 — Reflection Loop **What:** Agent generates → second prompt critiques its own output → agent revises. **Add when:** - First draft is almost-right but has a fixable, predictable flaw (wrong tone, missed edge case, minor hallucination) - The task is creative or open-ended (writing, code architecture, explanat