Video lesson coming soon
We're filming this one. The full written lesson below is ready to study right now.
Architecture starts with refusing complexity you don’t need. The first decision is the shape of the solution — and reaching for an agent when a rule would do is the most common AI-architecture mistake.
The pattern ladder
Rule / heuristic
deterministic, debuggable, free
Classic ML
patterns in tabular data, measurable
LLM call
language/understanding, one-shot
RAG
LLM grounded in your knowledge
Agent
multi-step, tools, loops — last resort
⚠
Every rung up adds failure modes
Each step up the ladder adds latency, cost, non-determinism, and new ways to fail. Climb only when the rung below genuinely can’t meet the requirement.
Match shape to requirement
| Requirement | Pattern |
|---|---|
| Deterministic policy, stable rules | rule engine |
| Predict a number/class from data | classic ML |
| Understand/transform language once | single LLM call |
| Answer from your private docs, cite-able | RAG |
| Multi-step task using tools + state | agent (carefully) |
Takeaway
Start at the bottom of the ladder and climb only when forced. Rules → ML → LLM → RAG → agent; each rung buys capability and pays in latency, cost and non-determinism.