Agentic Architectures
A unified Python library for running, comparing, and extending 35 agentic architecture patterns.
- Source repo
- FareedKhan-dev/all-agentic-architectures
- Stars
- ★ 4.7k
- Last updated
- 5mo ago
- License
- MIT
- Primary language
- Jupyter Notebook
- FA score
- 57/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformOpenAI API · Claude API
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A Python team wants to compare Reflection, Self-Consistency, and Tree of Thoughts while preserving one .run(task) integration point.
- Not a fit if
- Users who want a single fixed model and no multi-provider key setup
- Teams needing production browser automation: BrowserAgent requires Anthropic and real Playwright
- Source review
- 57/100 · Major gaps
What does this agent do, and when should you use it?
Agentic Architectures is a LangGraph-based Python library and runnable textbook with executed Jupyter notebooks. It packages 35 patterns as Architecture classes spanning reflection, search, RAG, memory, tool use, multi-agent coordination, and safety routing. Every architecture uses the same .run(task) interface and returns an ArchitectureResult; the example reads output and final_score metadata. It supports nine providers: Nebius, OpenAI, Anthropic, Groq, Ollama, Together, Fireworks, Mistral, and Google, selected through environment configuration. The repository also includes a 17-task comparative benchmark, 35 executed notebooks, and tests for library and notebook integrity.
After installation, get_llm() creates an LLM client and a class such as Reflection(llm=get_llm(), max_iterations=2, target_score=8) can be instantiated. Calling arch.run("Write a haiku about a glacier.") runs the selected architecture and produces result.output plus result.metadata["final_score"]. Depending on the class, the flow can generate, critique, and refine; vote across samples; search trees; retrieve and grade documents; store memory; invoke tools; or coordinate multiple agents. Thirteen architectures use the deterministic-picker pattern: the LLM commits to boolean or enum features and Python composes the decision signal; nine more are described as immune to this scorer issue by design. BrowserAgent uses real Playwright with a safety gate, SWE-Agent is a sandboxed filesystem agent, and Voyager executes reusable Python skills through real subprocesses.
- A Python team wants to compare Reflection, Self-Consistency, and Tree of Thoughts while preserving one .run(task) integration point.
- An engineer building retrieval QA needs to test Agentic RAG, CRAG, Self-RAG, Adaptive RAG, and GraphRAG as alternative retrieval shapes.
- An automation team evaluating browser-operation risk wants to study BrowserAgent’s safety gate; production Computer-Use requires Anthropic.
- A developer designing coordinated research or article workflows wants runnable examples of Multi-Agent, Blackboard, Debate, STORM, and Meta-Controller.
- A technical lead choosing an architecture for a task wants to inspect the included 17-task benchmark and per-task answer excerpts.
How do you install or deploy this agent?
From a fresh clone:
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 .envFill .env with NEBIUS_API_KEY and the relevant credentials for the selected provider. The runnable-example extras are also documented as: pip install "agentic-architectures[nebius,faiss,tavily]".
How do you use this agent?
Set LLM_PROVIDER and the matching provider key, then run:
from agentic_architectures import get_llm
from agentic_architectures.architectures import Reflectionarch = 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")Swap the Architecture class while keeping the downstream .run(task) call and ArchitectureResult shape. Run pytest -q for tests; set RUN_INTEGRATION=1 to enable real-LLM integration tests.
What are this agent's strengths and limitations?
- All 35 architectures share a .run(task) contract and ArchitectureResult shape, making pattern swaps straightforward.
- The documented provider matrix covers nine providers, and the README explicitly states that every architecture works with OpenAI.
- The 17-task benchmark supplies comparative results and answer excerpts rather than only conceptual descriptions.
- The deterministic-picker pattern separates categorical LLM judgments from Python’s final decision composition to address LLM-as-Scorer flat-band behavior.
- Most runs require a chosen LLM provider, network access, and its API credentials; local Ollama needs no key, but tool calling depends on the model.
- The most recent reported benchmark used Nebius Llama-3.3-70B, so its results are not evidence that outcomes reproduce across providers or models.
- Thirty-seven integration tests are environment-gated; a real-LLM path for each architecture runs only with RUN_INTEGRATION=1.
- Production Computer-Use for BrowserAgent explicitly requires Anthropic, and its browser execution depends on real Playwright.
How does this agent compare with similar options?
The repository supports internal pattern comparison: Agentic RAG, CRAG, Self-RAG, Adaptive RAG, and GraphRAG for retrieval; Multi-Agent, Blackboard, Debate, STORM, and Meta-Controller for coordination. Its 17-task leaderboard also records pattern-fit failures, including LATS on arithmetic and Debate plus Ensemble on the Sally trick.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Agentic Architectures This agent | 57 · Major gaps | Library / SDKFree + model costs | ★ 4.7k | 5mo ago | Jupyter Notebook | OpenAI API · Claude API |
| GenAI Agents Comprehensive Tutorial Collection | 57 · Major gaps | Library / SDKFree + model costs | ★ 25k | 6d ago | Jupyter Notebook | OpenAI API |
| Agent Craft | 52 · Major gaps | CLIFree + model costs | ★ 536 | 12d ago | Python | OpenAI API |
| CUGA | 71 · Some gaps | CLIFree + model costs | ★ 1.1k | 5d ago | Python | OpenAI API |
How does FollowAgents rate this agent?
Why each dimension lost points
The CI workflow declares contents: read, and pyproject makes providers, vector stores, and tools optional extras; the installation guide also distinguishes local Ollama from services needing API keys, supporting narrower choices. The security policy identifies browser, filesystem, subprocess, and secret risks, while the README mentions a BrowserAgent safety gate and a Dry-Run approval gate. The supplied material does not show their implementations, actual confirmation flows, sensitive-data retention rules, or rollback mechanisms, so those areas receive only partial credit. Dependencies have open-ended minimum-version constraints, with no lock or vulnerability-mitigation evidence shown. The README lists research and project references, aiding attribution, but the publisher identity remains unverified as specified.
The README and CI configuration describe test counts, architecture coverage, and environment-gated integration tests, indicating some validation planning; these are static claims and do not establish that runs succeeded. CI explicitly allows mypy failures, and the docs workflow allows a failed strict build to continue to deployment. Dependency ranges are broad, and the supplied material shows little about user-facing failure messages or recovery, so reliability scores remain limited.
The materials address learners, developers, and researchers with goal-based learning paths, architecture families, and benchmark tables. A shared Architecture.run interface and provider extras support switching, but provider capabilities and per-architecture fit are only partly described. BrowserAgent, web search, and vector backends have environment or service requirements; installation notes cover some, not every deployment context.
The README, installation and quickstart guides, architecture index, tutorials, and API documentation entry points form a clear information structure. Setup steps, extras, key configuration, and browser dependencies are concrete. Examples and benchmark material are extensive, but no FAQ is shown and limitations are scattered. The MIT LICENSE is explicit. Version evidence conflicts between 0.2.x in the security policy and 0.3.0 in pyproject, and the supplied packet does not include changelog contents for reconciliation. The security policy gives a reporting channel and response targets, providing some maintenance responsibility evidence, though publisher identity is unverified.
A shared interface, 35 patterns, runnable textbook framing, and task-oriented benchmark results could help users compare and compose architectures; the example also specifies consumable output, state, trace, and metadata fields. Claims such as production-grade readiness and cost advantages are mostly project assertions. The packet does not provide complete supporting benchmark data for the cited cost, duration, and outcomes, so actual benefit relative to operating cost is hard to judge.
The README's architecture tables and documentation links, pyproject metadata, CI configuration, and security policy provide traceable evidence for some claims. Files partly corroborate the package name, installation approach, and shared interface, and the test scope and integration gate are described concretely. However, README test and benchmark results are not independently verified here; version information conflicts, and the material does not consistently separate promotional conclusions from facts supported by the supplied evidence.
- The README's test counts, benchmark scores, and cost figures are static claims and were not executed or independently verified in this review.
- The security policy names risk areas in BrowserAgent, SWEAgent, and Voyager, but the supplied material does not show the corresponding control implementations; review those code paths before using real browser, filesystem, or subprocess actions.
- Version evidence conflicts: SECURITY.md calls 0.2.x current, while pyproject.toml declares 0.3.0.
Maintain this project? Add the FARS badge to your README
[](https://followagents.com/en/agents/all-agentic-architectures)<a href="https://followagents.com/en/agents/all-agentic-architectures"><img src="https://followagents.com/badge/all-agentic-architectures.svg" alt="FARS review"></a>Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
[0.3.0](https://github.com/FareedKhan-dev/all-agentic-architectures/compare/v0.2.0...v0.3.0) (2026-05-28) Features * 17 foundational architecture patterns ([c09a801](https://github.com/FareedKhan-dev/all-agentic-architectures/commit/c09a8012854069efe6a7d8f6b4a276be0056ca5d)) * 18 advanced architecture patterns ([af34630](https://github.com/FareedKhan-dev/all-agentic-architectures/commit/af3463027095ea114aa30775e86d6f6a4e63e9f9)) …View on GitHub ↗
FAQ
Is this a hosted chat product?
What cost and runtime should I expect?
Am I locked into Nebius?
Does it perform real browser, filesystem, or process actions?
Thanks — we use this to improve the page.