Skip to main content
Agentic Patterns gives you composable building blocks for constructing multi-step AI agent pipelines. Whether you need a simple plan-and-execute loop or a full LangGraph-powered state machine with automatic retry logic, you can assemble production-ready agentic systems from well-defined, interchangeable components.

Quickstart

Install dependencies and run your first agentic workflow in minutes.

Configuration

Configure LLMs, memory, agents, and tools for your environment.

​
What is Agentic Patterns?

Agentic Patterns is a Python library (agenticpatterns) that layers three architectural ideas on top of LangChain and LangGraph:
  1. Role-specific agents — discrete classes for planning, execution, monitoring, and local summarization.
  2. Workflow patterns — sequential and parallel orchestration strategies that coordinate agents.
  3. Supporting infrastructure — SQLite short-term memory with exact-match caching, context compression, and web search via curl.
All components are designed to work together or independently. You can drop a PlanningAgent into any LangChain chain, use SQLiteShortTermMemory for session persistence in a chatbot, or run the full LangGraphOrchestrator as a self-correcting pipeline.

​
Key architectural components

Planning agent

PlanningAgent breaks a high-level task into a structured JSON plan — a list of sequential steps passed to the executor.

Execution agent

ExecutionAgent carries out individual plan steps. It optionally wraps a LangGraph ReAct agent when tools such as CurlSearchTool are provided.

Monitoring agent

MonitoringAgent evaluates each step’s output against its original objective and returns structured JSON feedback with a success flag.

Local agent

LocalAgent is a lightweight summarization agent that aggressively condenses context to save tokens for downstream steps.

Sequential workflow

SequentialWorkflow runs planning → execution → monitoring in order, threading the output of each step as context into the next.

Parallel workflow

ParallelWorkflow executes independent tasks concurrently using a thread pool, returning all results when the batch completes.

LangGraph orchestrator

LangGraphOrchestrator builds a LangGraph StateGraph that coordinates planner, executor, and monitor nodes with configurable retry logic.

Short-term memory

SQLiteShortTermMemory persists session messages locally and provides exact-match cache lookups to skip redundant LLM calls.

CompressContextTool

Strips whitespace, removes common filler words, and truncates text locally — no LLM call required — before injecting context into prompts.

CurlSearchTool

Queries the Wikipedia OpenSearch API via curl and returns the top three snippets, giving agents access to factual web content.

​
Local and cloud LLM support

Agentic Patterns works with any BaseChatModel from LangChain. Two configurations are used out of the box:
  • Local (Ollama) — ChatOllama connects to a locally running Ollama server. Set LOCAL_MODEL and OLLAMA_HOST in your .env file.
  • Cloud (OpenAI-compatible) — ChatOpenAI with a custom base_url connects to any OpenAI-compatible endpoint such as Routeway or LLM API. Set SUMMARY_HOST, SUMMARY_MODEL, and SUMMARY_AGENT_API_KEY.
You can mix both: the main orchestration pipeline can use a fast local model while the monitoring agent uses a more capable cloud model.
Agents are LLM-agnostic. Any BaseChatModel instance — including Anthropic, Groq, or Mistral — can be passed to any agent constructor.

​
How it fits together

Task
  └─► PlanningAgent        → produces a list of steps
        └─► ExecutionAgent  → executes each step (optionally with tools)
              └─► MonitoringAgent → validates output; retries on failure
                    └─► SQLiteShortTermMemory → caches the final answer
The LangGraphOrchestrator encodes this loop as a LangGraph state machine. SequentialWorkflow implements the same pattern in plain Python. Both expose a single .run(task=...) method and return structured result dictionaries. Get started with the Quickstart.

Build docs developers (and LLMs) love

Get started for freeTalk to us