Core Philosophy

The most successful implementations use simple, composable patterns — not complex frameworks.

Workflow Patterns

1. Prompt chaining: Sequential LLM calls for decomposed tasks
2. Routing: Classify inputs to specialized downstream tasks
3. Parallelization: Multiple LLMs process simultaneously (sectioning or voting)
4. Orchestrator-workers: Central LLM dynamically distributes subtasks
5. Evaluator-optimizer: One LLM generates, another evaluates, iterating to improve

When to Use Agents

For open-ended problems where steps can't be predicted or hardcoded. Agents use tools, get environment feedback, and loop until completion.

Key Principles

1. Simplicity: Start with single LLM + retrieval + examples
2. Transparency: Show the agent's planning steps clearly
3. Tool design: Invest more time in tool optimization than prompts. Use clear names, descriptions, examples. Design tools to be hard to misuse (poka-yoke).

"Success isn't building the most complex system. It's building the right system for your needs."