⚡ LightAgentX

A lightweight, modular agentic AI framework built from scratch.
No LangChain. No CrewAI. No magic. Just clean Python.

What is LightAgentX?

LightAgentX is a ~1,200 line Python framework that rebuilds every core pattern from LangChain and CrewAI from first principles — LLM abstraction, memory, tools, planning, the agent loop, and multi-agent orchestration.

Every module is transparent, composable, and designed to teach you exactly what happens when an AI agent "thinks", "remembers", or "uses a tool".

lightagentx/
├── llm/          LLM abstraction (swap providers without changing agent code)
├── memory/       Conversation memory (buffer + LLM-compressed summary)
├── tools/        Tool system (@tool decorator → JSON Schema → execution)
├── planning/     ReAct reasoning (Thought → Action → Observation)
├── loop/         Core agent loop (the heart of ALL agentic frameworks)
├── agents/       Single agents, sequential pipelines, crew orchestration
└── utils/        Colored logger for tracing

Installation

pip install lightagentx

Or install from source:

git clone https://github.com/devansh/lightagentx.git
cd lightagentx
pip install -e .

Quick Start

from lightagentx import OpenAILLM, SingleAgent, tool

# 1. Create an LLM
llm = OpenAILLM()  # uses OPENAI_API_KEY env var

# 2. Define tools
@tool
def calculate(expression: str) -> str:
    """Evaluate a math expression.

    Args:
        expression: A math expression like '2 + 3 * 4'.
    """
    return str(eval(expression))

# 3. Create an agent
agent = SingleAgent(name="MathBot", llm=llm, tools=[calculate])

# 4. Run it
print(agent.run("What is 25 * 37?"))
# → "25 × 37 = 925"

📖 Full usage guide → docs/guide.md

Features

Examples

# Basic chat with memory
python examples/01_basic_chat.py

# Agent with tools (calculator, weather, time)
python examples/02_tool_agent.py

# ReAct reasoning loop
python examples/03_react_agent.py

# Multi-agent pipelines + crew
python examples/04_multi_agent.py

Testing

All tests use mocked LLM responses — no API key required:

pip install -e ".[dev]"
python -m pytest tests/ -v

72 passed in 0.96s ✅

How It Maps to LangChain & CrewAI

Architecture

Dependency flow (each layer only knows about layers below it):

  logger → llm → memory → tools → planning → loop/engine → agents

Design patterns used:

• Strategy — BaseLLM, BaseMemory, BasePlanner (swappable implementations)
• Decorator — @tool converts functions into tool objects
• Composite — Pipelines and crews are agents themselves (nestable)

License

MIT — see LICENSE.