// the find
FareedKhan-dev/all-agentic-architectures
35 production-grade agentic AI architectures (Reflexion, LATS, GraphRAG, MemGPT, Voyager, BrowserAgent, ...) — a Python library and runnable textbook with multi-provider LLM support and a 17-task benchmark leaderboard.
A Python library that wraps 35 agentic AI architectures (ReAct, LATS, MemGPT, STORM, GraphRAG, and more) behind a single `.run(task)` interface, paired with executed Jupyter notebooks that serve as a runnable textbook. It's for developers who want to understand or compare these patterns without reading five different papers and stitching together five different codebases.
The deterministic-picker discipline is a real contribution — having the LLM commit to booleans and enums rather than scoring on a flat 1-10 scale actually solves the variance problem in LLM-as-judge setups, and it's applied consistently across 13 architectures. The uniform `ArchitectureResult` return shape means you can genuinely swap Reflection for LATS with one line change and your downstream code doesn't care. The benchmark is honest: it reports group-think failures in Debate and Ensemble, wrong-shape failures in LATS on arithmetic, and doesn't hide the 22% miss rate. Test setup is clean — 283 fast unit tests that mock nothing live, env-gated integration tests behind `RUN_INTEGRATION=1` that hit real LLMs.
Everything is built on LangGraph, so you're importing that abstraction whether it fits your use case or not — if LangGraph's state machine model doesn't match what you're building, you'll fight the library rather than learn from it. The 'production-grade' label in the description overstates it: Voyager and BrowserAgent are production-adjacent, but Dry-Run, Constitutional AI, and several others are paper demonstrations with a thin wrapper. The benchmark runs on Llama-3.3-70B via Nebius and the results may not transfer to GPT-4o or Claude, especially for architectures like STORM where long-context model capability drives the outcome. Jupyter notebooks as the primary docs format ages poorly — outputs go stale, diffs are unreadable, and the notebook-execute CI workflow adds a meaningful maintenance burden every time an upstream API changes.