Pattern 1: Retrieval as a product surface
Treat the index, chunking strategy, and access policy as a product with owners — not a side effect of the prompt. When retrieval is wrong, generation will confidently amplify the mistake.
Pattern 2: Thin agents, thick tools
Prefer a small orchestration layer that calls well-defined tools (search, calculators, internal APIs) over a sprawling multi-agent graph you cannot debug. Add agents when a workflow truly needs them.
Pattern 3: Evaluation before ornamentation
UI polish and model upgrades feel productive. An offline eval set of real user questions feels boring — and is usually the difference between a launch and a rollback.
Pattern 4: Cost and latency as architecture
Caching, routing (small vs large models), and batching are design decisions. If you only discover them in the cloud bill, the architecture was incomplete.
What to avoid
- One mega-prompt that encodes policy, tools, and tone.
- Framework lock-in before you understand your failure modes.
- Shipping without an owner for quality after week one.