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.
Last verified:
Visit All Agentic Architectures
What is All Agentic Architectures?
All Agentic Architectures is a Python library and living textbook that packages 35 production-grade agentic AI patterns from the literature as runnable Architecture classes with a uniform contract. Each pattern ships with a fully executed Jupyter notebook whose theory is written against captured runs, not synthetic examples. The library supports 9 LLM providers (Nebius, OpenAI, Anthropic, Groq, Ollama, Together, Fireworks, Mistral, Google) and is built on top of LangGraph state machines.
Key features include the deterministic-picker pattern applied in 13 of 35 architectures, a 17-task benchmark suite with a comparative leaderboard ranking every architecture against relevant tasks, 283 passing tests, and a uniform .run(task) interface across all architectures. The repository covers architecture families including Reasoning & Reflection, Sampling & Search, Retrieval (RAG), Memory, Tools & Actions, Multi-Agent systems, Safety & Routing, and Specialty Patterns.
This tool is designed for AI developers, researchers, and students who want to understand and implement agent-based systems. It bridges the gap between theoretical research on autonomous agents and real software implementations. Users can swap architecture patterns without changing downstream code, making it ideal for prototyping, production deployment, and educational purposes.
The library includes practical implementations like BrowserAgent with real Playwright integration, SWE-Agent for sandboxed file-system operations, MemGPT for OS-style context tiers, and GraphRAG with knowledge graphs and community summaries. It features MIT licensing and active community contributions.
All Agentic Architectures pricing
Pricing model: Freemium
Free and open-source under MIT license. The library is completely free with no paid tiers. Users only pay for LLM API usage through their chosen provider (Nebius default, OpenAI, Anthropic, Groq, Ollama, Together, Fireworks, Mistral, or Google). The included demos use Nebius Llama-3.3-70B which is noted as cheapest, with benchmark runs costing approximately $1.50 in tokens for ~25 minutes of execution.
All Agentic Architectures pros
- 35 production-grade agentic AI patterns in one library
- Uniform .run(task) interface across all architectures
- 283 passing tests with 37 skipped integration tests
- 17-task benchmark suite with comparative leaderboard
- Supports 9 LLM providers without code changes
- Fully executed Jupyter notebooks with real LLM outputs
- Deterministic-picker pattern escapes LLM-as-Scorer pathology
- Built on LangGraph state machines for reliability
- MIT licensed open-source repository
- Multi-provider support with simple env var switching
- Real Playwright integration for BrowserAgent
- 3.4k GitHub stars and 594 forks showing community trust
- Detailed documentation with dark-mode MkDocs site
- Pre-built learning paths for different use cases
- Docker and devcontainer support for easy setup
All Agentic Architectures cons
- Requires API keys for most LLM providers
- Integration tests gated by RUN_INTEGRATION env variable
- Tool calling depends on model for Ollama local usage
- Some architectures fail on specific task types
- LATS performs poorly on arithmetic tasks
- Debate and Ensemble show group-think on trick questions
- Reflexion and AWM fail on raw fact recall tasks
- Requires virtualenv setup with multiple extras
Frequently asked questions about All Agentic Architectures
What is the deterministic-picker pattern?
The deterministic-picker pattern is the central technical discipline of the repository. Every LLM-as-Scorer surface has the LLM commit to categorical features (booleans, enums) and lets Python compose the deciding signal. This is the universal escape from the LLM-as-Scorer flat-band pathology, applied in 13 of 35 architectures, with 9 more being architecturally immune by design.
Can I switch between LLM providers?
Yes, you can switch across 9 providers (Nebius, OpenAI, Anthropic, Groq, Ollama, Together, Fireworks, Mistral, Google) by setting the LLM_PROVIDER environment variable and filling in the corresponding key in your .env file. No code changes are required to swap providers.
What architecture families are included?
The library includes 7 architecture families: Reasoning & Reflection (Reflection, Reflexion, Chain-of-Verification, Self-Discover, Constitutional AI), Sampling & Search (Self-Consistency, Tree of Thoughts, LATS, Mental Loop, Ensemble), Retrieval/RAG (Agentic RAG, CRAG, Self-RAG, Adaptive RAG, GraphRAG), Memory (Episodic+Semantic, Graph Memory, MemGPT, Voyager, Agent Workflow Memory), Tools & Actions (Tool Use, ReAct, Planning, PEV, SWE-Agent, BrowserAgent), Multi-Agent (Multi-Agent, Blackboard, Debate, STORM, Meta-Controller), and Safety/Routing/Specialty patterns (Dry-Run, Reflexive Metacognitive, RLHF Self-Improvement, Cellular Automata).
What is the benchmark performance?
The benchmark suite has 17 tasks and achieves 33/42 correct (78% overall). Strong performers include Reflection, SelfConsistency, SelfDiscover, and BrowserAgent (2/2 or 3/3). Twenty-one more architectures achieved perfect scores on attempted tasks. Pattern-fit failures include LATS on arithmetic (wrong shape), Debate and Ensemble on Sally trick (group-think), and Reflexion with AWM on raw-fact recall (wrong memory shape).
What memory types are available?
The library includes 7 memory shapes: Episodic + Semantic (conversation turns + triples), Graph Memory (subject, predicate, object triples), MemGPT (OS-style context + archival tiers), Voyager (reusable Python skills via real subprocess), and Agent Workflow Memory (high-level workflow recipes). Each architecture uses the memory shape appropriate for its pattern.
How do I contribute a new architecture?
Follow the 5-step contributor recipe in CONTRIBUTING.md. The PR template includes a deterministic-picker checklist. For existing architecture improvements (bug fixes, prompt tuning, performance, scoring rubric), open an issue first to discuss scope. The project uses Conventional Commits and release-please auto-generates the CHANGELOG.
What providers are best for different use cases?
Nebius is default and cheapest for included demos (Llama-3.3-70B + Qwen3-Thinking). OpenAI has highest quality for reasoning patterns. Anthropic has strong long context and is required for production Computer Use. Groq offers fast inference for high-volume Self-Consistency. Ollama works locally with no API key but tool calling depends on model. Mistral provides EU-hosted option. Google offers Gemini 2.x via Generative AI API.
What learning path should I follow?
Four curated paths exist: Beginner (mental model) - Reflection → Tool Use → ReAct → Planning → Self-Consistency; RAG-focused (production retrieval) - Agentic RAG → CRAG → Self-RAG → Adaptive RAG → GraphRAG; Multi-agent (coordination) - Multi-Agent → Blackboard → Debate → STORM → Meta-Controller; Safety (guardrails) - Dry-Run → Constitutional AI → Reflexive Metacognitive → BrowserAgent with safety gate.