Agentic Architectures智能体架构

Thirty-five production-grade agentic AI patterns. End to end.三十五种生产级智能体 AI 模式，端到端实现。

A library and a living textbook — real LLM outputs, provider-agnostic, deterministic-picker discipline throughout, and a comparative benchmark leaderboard that ranks every architecture against every relevant task.这是一个库，也是一本动态教科书——包含真实的 LLM 输出、与提供商无关的架构、贯穿始终的确定性选择器（deterministic-picker）准则，以及一个将每种架构与相关任务进行对比的基准测试排行榜。

35   ARCHITECTURES

283   PASSING TESTS

17   BENCHMARK TASKS

9   LLM PROVIDERS

0   MOCKED RUNS

---

Overview概述

A single Python library that packages every major agentic AI pattern from the literature as a runnable Architecture class with a uniform contract. Each pattern ships with a fully executed Jupyter notebook whose theory is written against the captured run — not synthetic examples. The library is multi-provider (Nebius, OpenAI, Anthropic, Groq, Ollama, Together, Fireworks, Mistral, Google) and built on top of LangGraph state machines.这是一个单一的 Python 库，它将文献中所有主要的智能体 AI 模式封装为具有统一接口的可运行 Architecture 类。每种模式都附带一个完整执行的 Jupyter notebook，其理论基于实际运行结果而非合成示例编写。该库支持多家提供商（Nebius、OpenAI、Anthropic、Groq、Ollama、Together、Fireworks、Mistral、Google），并基于 LangGraph 状态机构建。

The central technical discipline of the repository is the deterministic-picker pattern — 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; 9 more are architecturally immune by design.本仓库的核心技术准则为“确定性选择器”模式——每个作为评分器的 LLM 界面都会让 LLM 提交分类特征（布尔值、枚举），并由 Python 代码组合出最终决策信号。这是规避“LLM 作为评分器”时出现性能平坦化病态问题的通用方案，已应用于 35 种架构中的 13 种；另外 9 种架构在设计上天然免疫此类问题。

---

Quickstart快速入门

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

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")

Same .run(task) interface across all 35 architectures. Same ArchitectureResult return shape. Swap the class, swap the pattern — your downstream code does not change.所有 35 种架构均使用相同的 .run(task) 接口，并返回统一的 ArchitectureResult 格式。只需更换类名即可切换模式，无需修改下游代码。

Set up a virtualenv from a fresh clone从全新的克隆库中设置 virtualenv

git clone https://github.com/FareedKhan-dev/all-agentic-architectures
cd all-agentic-architectures

python -m venv .venv
.venv\Scripts\activate              # Windows
source .venv/bin/activate           # macOS / Linux

pip install -e ".[dev,test,docs,nebius,faiss,tavily,networkx]"
cp .env.example .env                # then fill in NEBIUS_API_KEY etc.

pytest -q                           # 283 tests pass in ~30s

---

Architecture families架构系列

Reasoning & Reflection Self-critique loops that drive answer quality up through iteration. Reflection · Reflexion · Chain-of-Verification · Self-Discover · Constitutional AI	Sampling & Search Sample many paths or grow a tree with rewards. Self-Consistency · Tree of Thoughts · LATS · Mental Loop · Ensemble	Retrieval (RAG) Ground every claim — five retrieval shapes. Agentic RAG · Corrective RAG · Self-RAG · Adaptive RAG · GraphRAG
Memory Learn across calls — pick the storage shape. Episodic + Semantic · Graph Memory · MemGPT · Voyager · Agent Workflow Memory	Tools & Actions From one search tool to a real Chromium browser. Tool Use · ReAct · Planning · PEV · SWE-Agent · BrowserAgent	Multi-Agent Specialists, debate, multi-perspective research. Multi-Agent · Blackboard · Debate · STORM · Meta-Controller
Safety & Routing Categorical actions through deterministic Python gates. Dry-Run · Reflexive Metacognitive · Computer Use	Specialty Patterns with a unique shape. RLHF Self-Improvement · Cellular Automata	Cross-cutting Patterns that appear across families. Deterministic-picker · Memory variants

---

The 35 architectures35 种架构

Reasoning & Reflection推理与反思

Architecture	Pattern	Reference
Reflection	Generate → critique → refine	Madaan 2023
Reflexion	Verbal reflections in episodic memory	Shinn 2023
Chain-of-Verification (CoVe)	Verify each baseline claim independently	Dhuliawala 2023
Self-Discover	SELECT → ADAPT → IMPLEMENT → SOLVE	Zhou 2024
Constitutional AI	Per-rule pass/fail → revise	Bai 2022

Sampling & Search采样与搜索

Architecture	Pattern	Reference
Self-Consistency	Sample N paths, majority-vote	Wang 2022
Tree of Thoughts	Beam search over thoughts	Yao 2023
LATS	MCTS tree with reward backup	Zhou 2024
Mental Loop	Simulate → score (deterministic-picker)	this repo
Ensemble	N voters, weighted aggregation	this repo

Retrieval (RAG)检索 (RAG)

Architecture	Pattern	Reference
Agentic RAG	Agent decides when & what to retrieve	LangGraph reference
Corrective RAG (CRAG)	Grade docs, fall back to web	Yan 2024
Self-RAG	Per-doc reflection tokens	Asai 2024
Adaptive RAG	Pre-route by query complexity	Jeong 2024
GraphRAG	KG + community summaries	Microsoft 2024

Memory记忆

Architecture	Stored unit	Reference
Episodic + Semantic	Conversation turns + triples	Park 2023
Graph Memory	(subject, predicate, object) triples	this repo
MemGPT	OS-style context + archival tiers	Packer 2023
Voyager	Reusable Python skills (real subprocess)	Wang 2023
Agent Workflow Memory	High-level workflow recipes	Wang 2024

Tools & Actions工具与行动

Architecture	Pattern	Reference
Tool Use	Agent with one tool	LangChain reference
ReAct	Thought → Action → Observation	Yao 2022
Planning	Decompose → execute → replan	Wei 2022
Plan-Execute-Verify (PEV)	Post-execution verification per step	this repo
SWE-Agent	Sandboxed file-system agent	Yang 2024
BrowserAgent	Real Playwright + safety gate	Anthropic Computer-Use 2024

Multi-Agent多智能体

Architecture	Pattern	Reference
Multi-Agent	Supervisor + specialists	LangGraph reference
Blackboard	Shared workspace + agents	classical AI
Debate	N agents × K rounds	Du 2023
STORM	Multi-perspective research → article	Shao 2024
Meta-Controller	Router over architectures	this repo

Safety, Routing & Specialty安全、路由与专业领域

Architecture	Pattern	Reference
Dry-Run	Propose → simulate → approval gate	this repo
Reflexive Metacognitive	Self-aware capability routing	this repo
RLHF Self-Improvement	Multi-dim deterministic scoring + archive	this repo
Cellular Automata	LLM rules over a grid	this repo

---

Provider compatibility提供商兼容性

Provider	Install extra	Notes
Nebius  (default)	[nebius]	Llama-3.3-70B + Qwen3-Thinking; cheapest for the included demos
OpenAI	[openai]	All architectures work; highest quality for reasoning patterns
Anthropic	[anthropic]	Strong on long context; required for production Computer-Use
Groq	[groq]	Fast inference; great for high-volume Self-Consistency
Ollama  (local)	[ollama]	No API key; tool calling depends on the model
Together	[together]	Wide model catalogue
Fireworks	[fireworks]	Function-calling first-class
Mistral	[mistral]	EU-hosted option
Google	[google]	Gemini 2.x via Generative AI API

Switch via LLM_PROVIDER + the corresponding key in .env. No code changes.通过 LLM_PROVIDER 和 .env 文件中的对应密钥进行切换，无需更改代码。

---

Benchmarks基准测试

A 17-task suite runs every architecture and scores results. Most recent run, real Nebius Llama-3.3-70B, ~25 min, ~$1.50 in tokens:包含 17 个任务的测试套件会运行每种架构并对结果进行评分。最近一次运行使用真实的 Nebius Llama-3.3-70B，耗时约 25 分钟，Token 成本约 1.50 美元：

Outcome	Architectures
Strong  2/2 or 3/3	Reflection  SelfConsistency  SelfDiscover  BrowserAgent
Perfect on attempted  1/1	21 more — see leaderboard
Pattern-fit failures	LATS on arithmetic (wrong shape) · Debate + Ensemble on Sally trick (group-think) · Reflexion + AWM on raw-fact recall (wrong memory shape)
Overall	33 / 42 correct  78%

_{Full leaderboard with per-task answer excerpts: docs/benchmarks.md}包含各任务答案摘要的完整排行榜：docs/benchmarks.md

---

Learning paths学习路径

Four curated reading orders, depending on what you're trying to do.根据你的目标，提供四种精选阅读顺序。

Path	For	Order
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 (safety gate)

---

Star historyStar 历史

---

Tested已测试

pytest -q
283 passed, 37 skipped (env-gated integration), 1 warning in ~30s

Suite	Coverage
Registry sweep	All 35 architectures (metadata + instantiate + build)
Pure-Python helpers	Haiku checker, composite scorers, subprocess executor, safety gate, sandbox path
Notebook integrity	All 35 notebooks executed, no error outputs, §9 commentary tailored from real captured runs
Integration  (env-gated)	One real-LLM happy-path per architecture, gated via RUN_INTEGRATION=1

---

Documentation文档

Full docs site	Dark-mode site with embedded notebooks  (live after first deploy)
Quickstart	One-command install, 8-line example
Switching providers	Capability matrix; one env var to swap
Add your own architecture	5-step contributor recipe
Deterministic-picker pattern	The central technical pattern, explained once
Memory variants	Comparison of all 7 memory shapes
API reference	mkdocstrings auto-gen from docstrings  (live after first deploy)
Benchmarks	Full per-task leaderboard with answer excerpts

---

Contributing贡献

Contributions welcome. Two paths:欢迎贡献。有两种途径：

1. Add a new architecture — follow the 5-step recipe. The PR template includes a deterministic-picker checklist.添加新架构——遵循 5 步指南。PR 模板包含确定性选择器的检查清单。
2. Improve an existing one — bug fix, prompt tuning, performance, scoring rubric. Open an issue first to discuss scope.改进现有架构——包括 Bug 修复、提示词优化、性能提升或评分准则优化。请先开启 Issue 讨论范围。

See CONTRIBUTING.md for the dev setup, code style, and commit-message convention (Conventional Commits — release-please auto-generates the CHANGELOG).请参阅 CONTRIBUTING.md 了解开发环境设置、代码规范和提交信息约定（Conventional Commits——release-please 会自动生成 CHANGELOG）。

---

Citation引用

@misc{khan2026agentic,
  title         = {Agentic Architectures: A Library of 35 Production-Grade Agentic AI Patterns},
  author        = {Khan, Fareed},
  year          = {2026},
  howpublished  = {\url{https://github.com/FareedKhan-dev/all-agentic-architectures}},
  note          = {MIT licensed Python library and runnable textbook}
}

---

License许可证

MIT — © 2026 Fareed Khan.MIT — © 2026 Fareed Khan。

_{Built on LangGraphLangGraph  ·  Docs powered by Material for MkDocsMaterial for MkDocs  ·  Default LLM via NebiusNebius 基于 LangGraph 构建  ·  文档由 Material for MkDocs 提供支持  ·  默认 LLM 由 Nebius 提供}

_{★ Star★ Star  ·  ForkFork  ·  Open an issue开启 Issue  ·  Discuss讨论 ★ Star  ·  Fork  ·  开启 Issue  ·  讨论}