Architecture & Patterns: LLM Fundamentals for Agents¶

Production Patterns¶

Architecture View¶

graph LR
  A[Task input] --> B[Prompt template]
  B --> C[Context budget check]
  C --> D[Model routing]
  D --> E[Scored output]

Common Failure Modes¶

Prompt templates that mix instructions and examples inconsistently.
Ignoring output-token costs while optimizing only input size.
Routing rules too coarse for mixed-complexity tasks.

Interview Q&A¶

Q: Interview Q: Why not always use the strongest model?

Strong models improve difficult tasks, but universal use creates unnecessary cost and latency. Routing preserves quality where needed and lowers operational spend elsewhere.

Q: Interview Q: How do you reduce lost-in-the-middle behavior?

Move key constraints earlier, trim irrelevant context, and explicitly restate critical facts near the end before generation.