The complete library of agentic AI patterns.¶
Thirty-five production-grade architectures, built on LangGraph. Real LLM outputs end-to-end, provider-agnostic, deterministic-picker discipline throughout — and a benchmark that ranks them all.
Living textbook¶
Every notebook's commentary quotes the actual captured run — not a synthetic example. Theory written against observed behavior.
Deterministic picker¶
Every LLM-as-Scorer surface uses a categorical commitment + Python composition. No flat-band pathology, by design.
Provider agnostic¶
One get_llm() factory speaks Nebius, OpenAI, Anthropic, Groq, Ollama, Together, Fireworks, Mistral, Google.
Comparable¶
A 17-task benchmark suite runs every architecture and scores results. See which pattern suits which task family.
Eight families. Thirty-five patterns.¶
Reflect, verify, revise.
Self-critique loops that drive answer quality up through iteration. Catch hallucinations before they ship.Reflection · Reflexion · Chain-of-Verification · Self-Discover · Constitutional AIMany paths, one answer.
Sample N reasoning trajectories and pick the modal answer; or grow a tree and search it with rewards.Self-Consistency · Tree of Thoughts · LATS · Mental LoopGround every claim.
Five distinct retrieval shapes — from agent-decides-when, to corrective grading, to graph community summaries.Agentic RAG · Corrective RAG · Self-RAG · Adaptive RAG · GraphRAGLearn across calls.
Episodic reflections, archival skills, OS-style tiered context, mined workflows — pick the right shape for your task.Episodic + Semantic · MemGPT · Voyager · Agent Workflow MemoryAct on the world.
From a single search tool to a sandboxed code repo to a real Chromium browser.Tool Use · ReAct · Planning · PEV · SWE-Agent · BrowserAgentQuickstart¶
from agentic_architectures import get_llm
from agentic_architectures.architectures import Reflection
arch = Reflection(llm=get_llm())
result = arch.run("Write a haiku about the sea.")
print(result.output)
print("score:", result.metadata["final_score"], "/ 10")
Same .run(task) interface across all 35 architectures. Same ArchitectureResult shape. Swap the class, swap the pattern — your downstream code doesn't change.