All-Agentic-Architectures is a Python library that implements production-grade agentic AI patterns as a runnable textbook with multi-provider LLM support and a benchmark leaderboard.
The library solves the problem of fragmented agentic AI implementations by packaging major patterns from the literature—including Reflexion, LATS, GraphRAG, MemGPT, Voyager, and BrowserAgent—into a uniform interface. Each architecture is implemented as a class with the same `.run(task)` contract and `ArchitectureResult` return shape, allowing developers to swap patterns without changing downstream code. The implementations are built on LangGraph state machines and run against real LLM outputs rather than synthetic examples. A core technical discipline is the deterministic-picker pattern, which addresses the LLM-as-Scorer flat-band problem by having the LLM commit to categorical features and letting Python compose the deciding signal; this pattern is applied across multiple architectures.
Developers should choose this tool if they need to experiment with or deploy multiple agentic patterns within a single codebase, or if they want to understand how different architectures perform on comparable tasks. The library supports nine LLM providers including OpenAI, Anthropic, Groq, Ollama, Together, Fireworks, Mistral, Google, and Nebius, making it suitable for projects that need provider flexibility. The benchmark leaderboard ranks all architectures against a set of tasks, enabling comparative evaluation. The project is structured as both a library and a living textbook, with each pattern accompanied by a fully executed Jupyter notebook whose theory is grounded in captured runs.
Development activity shows consistent focus on test coverage and deterministic behavior. The repository maintains passing tests across its implementations and emphasizes zero mocked runs, ensuring all examples reflect actual LLM behavior. The codebase demonstrates architectural discipline through systematic application of the deterministic-picker pattern to eliminate scoring ambiguity.
