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

Summary

all-agentic-architectures is a Python library and runnable textbook that implements 35 agentic AI patterns behind a common Architecture class, .run(task) method, and ArchitectureResult return shape. It also provides executed Jupyter notebooks, LangGraph-based implementations, nine LLM providers, and a 17-task comparison benchmark. Its central design rule, the “deterministic-picker,” asks an LLM for categorical decisions such as booleans or enums and lets Python combine them.

What it is useful for

The library is useful for comparing agent designs without rewriting downstream code. The README groups the patterns into reasoning and reflection, sampling and search, retrieval, memory, tools and actions, multi-agent coordination, and safety or routing. For example, an engineer could begin with Reflection to generate, critique, and refine an answer, then compare it with GraphRAG for knowledge-grounded retrieval or BrowserAgent, which uses Playwright and a safety gate, without changing the calling interface.

It also serves as a study and evaluation reference. Each architecture has a notebook written around captured runs, and the benchmark leaderboard reports results per task. The latest reported run, using Nebius Llama-3.3-70B, got 33 of 42 answers correct; the README also records failures such as LATS on arithmetic, group-think in Debate and Ensemble, and weak raw-fact recall with Reflexion and Agent Workflow Memory. Those are documented pattern-fit limitations, not evidence that every pattern suits every task.

Provider support is explicit: Nebius, OpenAI, Anthropic, Groq, Ollama, Together, Fireworks, Mistral, and Google. Documented caveats include Ollama tool calling depending on the model and Anthropic being required for production Computer-Use.

How engineers can use it

For a packaged install, the README documents:

pip install "agentic-architectures[nebius,faiss,tavily]"

Then configure LLM_PROVIDER and the matching key in .env, and use the common API:

from agentic_architectures import get_llm
from agentic_architectures.architectures import Reflection

arch = Reflection(llm=get_llm(), max_iterations=2, target_score=8)
result = arch.run("Write a haiku about a glacier.")

print(result.output)
print("score:", result.metadata["final_score"], "/ 10")

For development from a clone, the README also documents creating a virtual environment, running pip install -e ".[dev,test,docs,nebius,faiss,tavily,networkx]", copying .env.example to .env, and filling in NEBIUS_API_KEY or another provider key. It lists pytest -q as the test command.

A practical starting point is to read the matching notebook, select the architecture whose behavior fits the task, and use the same .run(task) call in the application. For setup details beyond these commands, consult the repository’s quickstart and provider documentation.