35 ways to make an LLM agent, actually implemented

What it does

Packages 35 named agentic architectures — from ReAct and Reflexion to GraphRAG and MemGPT — as interchangeable Python classes with a uniform .run(task) interface. Each ships as a fully executed Jupyter notebook where the explanation is written against captured LLM output, not synthetic demos. Built on LangGraph state machines and supports nine providers (Nebius, OpenAI, Anthropic, Groq, Ollama, Together, Fireworks, Mistral, Google).

The interesting bit

The “deterministic-picker” discipline: instead of asking an LLM to score its own work on a 1–10 scale (which tends to collapse into a narrow band), the LLM emits categorical features — booleans, enums — and Python composes the decision signal. The repo applies this escape hatch to 13 architectures and notes where 9 others are immune by design. It’s a rare case where the methodology section is the selling point.

Key highlights

• 283 passing tests, 17 benchmark tasks, zero mocked runs — every notebook shows real LLM output
• Provider-agnostic get_llm() wrapper; swap Nebius for Ollama without touching architecture code
• Families span reasoning, search, RAG, memory, tools, multi-agent, safety, and oddities like Cellular Automata
• Includes comparative benchmark leaderboard ranking architectures against relevant tasks
• Meta-Controller architecture routes queries across the other 34 implementations

Caveats

• README truncates provider compatibility table mid-row; full matrix lives elsewhere
• “Production-grade” is the repo’s claim — actual production readiness for each pattern is unclear
• Default provider is Nebius, which may require API key setup not all readers will have

Verdict

Worth bookmarking if you’re building agents and tired of reading papers that never ship code. Skip if you need a single polished framework rather than a comparative survey; this is a textbook with batteries included, not an opinionated product.