I’ve been building production LLM systems for a few years and kept hitting the same frustration: the conceptual literature (Ng, ReAct, ToT, Agentic RAG survey) is solid, but the gap between “here’s the pattern” and working, runnable code that handles real edge cases is enormous.
Built a reference repo to close that gap: GitHub - gtesei/agentic_design_patterns: Agentic AI Design Patterns · GitHub
29 patterns across foundational (prompt chaining, routing, reflection, RAG, ReAct, HITL…), advanced reasoning (Tree/Graph of Thoughts, deep research), reliability, orchestration, observability, and memory/learning. Each has runnable examples and documented tradeoffs.
Would love feedback on: pattern gaps, taxonomy, and whether quantitative tradeoff documentation (latency, token cost, quality delta) is the right frame or missing dimensions.
MIT-licensed, actively developing.