all-agentic-architectures: a library of 35 production-grade agentic AI patterns with a unified API and comparative benchmarks

What it solves

This project provides a unified Python library and educational resource for implementing agentic AI patterns. It eliminates the need to manually implement complex agent architectures from literature, providing 35 production-grade patterns with a consistent API, making it easy to swap and compare different agentic strategies for a given task.

How it works

Built on top of LangGraph state machines, the library packages each agentic pattern as a runnable Architecture class. It uses a "deterministic-picker pattern" to ensure that LLM-as-Scorer surfaces commit to categorical features (like booleans or enums) rather than free-text, allowing Python to handle the final decision logic. The library is provider-agnostic, supporting multiple LLM providers including Nebius, OpenAI, Anthropic, Groq, and others via a single environment variable.

Who it’s for

AI developers and researchers who want to implement, test, and benchmark different agentic architectures (such as Reflection, ReAct, or GraphRAG) without writing boilerplate code, as well as students of agentic AI who want a "runnable textbook" of these patterns.

Highlights

• 35 Production-Grade Architectures: Covers Reasoning & Reflection, Sampling & Search, RAG, Memory, Tools & Actions, and Multi-Agent systems.
• Deterministic-Picker Discipline: A technical approach to avoid "flat-band pathology" in LLM scoring by using categorical commitments.
• Multi-Provider Support: Seamlessly switch between 9 different LLM providers without changing code.
• Comparative Benchmarks: Includes a 17-task suite to rank architectures against specific tasks.
• Real-World Evidence: Every pattern includes a Jupyter notebook with theory written against actual captured LLM outputs.