agentik-framework 0.1.2

Lightweight, Modular Agent Framework for AI planning, tools, and memory

🤖 Agentik – Modular Agentic AI Framework for Python

A lightweight, modern, and pluggable Python framework to build AI agents with reasoning, memory, tool-use, and LLM backend support — designed to be minimal, extensible, and developer-friendly.

---

🔍 What is Agentik?

Agentik is a Python library that helps you build intelligent agents that can reason, plan, act, and reflect using any large language model (LLM) backend (e.g., OpenAI, Claude, Mistral via OpenRouter, or your local model). It abstracts the complexity and gives you:

• 🔌 LLM plugin support (OpenAI, Claude, etc.)
• 🧠 Memory integration (Dict, JSON, Vector optional)
• 🛠 Tool support (web search, calculator, file I/O)
• ⚙️ YAML or Python-based config
• 🖥 CLI for easy launching and debugging
• 🔒 API-key agnostic — the user supplies their own keys

---

🚀 Features

• Define agents in Python or YAML
• Core agent loop: plan → act → reflect → iterate
• Pluggable tools and memory systems
• Easily extend tools with Python
• Support for streaming, retry, and debugging
• CLI interface built with Typer
• Designed to be lightweight, transparent, and powerful

---

📁 Folder Structure

agentik/
│
├── agent.py          # 🧠 Core agent class and reasoning loop
├── llms.py           # 🔌 LLM backend interfaces (OpenAI, Claude, etc.)
├── tools/            # 🛠 Built-in and user-defined tools
│   └── __init__.py
│   └── calculator.py
│   └── websearch.py
│   └── filereader.py
│   └── ...
├── memory.py         # 🧠 Memory backends (Dict, JSON, FAISS, etc.)
├── config.py         # 🧾 YAML/JSON config parser using pydantic
├── cli.py            # 💻 CLI interface using Typer
├── utils.py          # ⚙️ Logger, token counter, helper functions
│
examples/             # 🎯 Sample agents and use-cases
tests/                # 🧪 Unit and integration tests
configs/              # ⚙️ YAML-based agent configuration files

---

🧠 Agent Workflow

The typical execution flow of an Agent in agentik:

[User Prompt] 
    ↓
[Agent.run(prompt)]
    ↓
[Plan Step] → Uses LLM to decide next action/tool
    ↓
[Act Step] → Runs selected tool or memory query
    ↓
[Reflect Step] → Evaluates outcome and updates memory
    ↓
[Repeat or Finish]

---

📦 Installation

Python 3.10+ is required

pip install agentik-framework

---

🛠 Usage

✨ Create an Agent (Python Way)

from agentik import Agent
from agentik.tools import WebSearchTool, CalculatorTool
from agentik.llms import OpenAIModel
from agentik.memory import JSONMemoryStore

agent = Agent(
    name="AIAssistant",
    goal="Find and summarize the latest AI trends.",
    tools=[WebSearchTool(), CalculatorTool()],
    llm=OpenAIModel(api_key="your-api-key-here", model="gpt-4"),
    memory=JSONMemoryStore("memory.json")
)

agent.run("What's the latest in AI agent frameworks?")

---

⚙️ Create an Agent via YAML (config.yaml)

name: "ResearchBot"
goal: "Summarize recent advancements in generative AI"
llm:
  type: openai
  model: gpt-4
  api_key: "YOUR_KEY_HERE"
tools:
  - web_search
  - file_reader
memory:
  type: json
  path: "memory.json"

---

📂 Main File to Run

There are two main entrypoints to run the framework:

1. Python File

You can create and run a Python file such as run_agent.py:

from agentik.config import load_agent_config
from agentik.agent import Agent

agent = load_agent_config("configs/my_agent.yaml")
agent.run("What are LLM agents?")

2. CLI Command (Recommended)

agentik run configs/my_agent.yaml --verbose

Make sure you run this inside the directory containing your YAML and memory files.

---

🧠 Supported LLMs

agentik supports the following backends (via your API keys):

• OpenAI (via OpenRouter)
• Mistral (via OpenRouter)
• Claude (via OpenRouter)
• DeepSeek (via OpenRouter)
• Local models via REST API

All models are initialized via the llms.py interfaces and passed as parameters. No key is embedded in the code.

---

🔌 Built-in Tools

Tool Name	Description
CalculatorTool	Evaluate math expressions
WebSearchTool	Search web results using DuckDuckGo/SerpAPI
FileReaderTool	Read content from .txt or .md files
JsonTool	Work with structured JSON documents

Tools are extendable: just subclass Tool and implement a run(input) method.

---

📂 Where to Put Code?

File	What to write here
agent.py	Core agent planning loop (already built-in)
llms.py	Add your own LLM wrapper classes (if needed)
memory.py	Custom memory strategies (default: Dict, JSON)
tools/	Add new tools (e.g., mytool.py) and register dynamically
config.py	Modify or validate YAML loading logic (via pydantic)
cli.py	Main command-line logic (already pre-written via typer)
examples/	Place your sample agents or demo scripts
configs/	Store reusable YAML configuration for agents

---

🧪 Testing

pytest tests/

You can mock LLM API responses and test agent planning, memory storage, and tool execution.

---

🖥 CLI Commands

agentik run configs/my_agent.yaml
agentik list-tools
agentik create-agent
agentik explain memory

Flags:

• --verbose → detailed logs
• --dry-run → simulate without actual LLM/API calls

---

📚 Documentation

• Markdown-based, hosted on GitHub Pages
• Located in docs/ (optional)
• Includes: Getting Started, API Reference, Tool Dev Guide

---

🧑‍💻 Contributing

We welcome contributions!

• Fork this repo
• Add your tool or feature
• Write tests in tests/
• Submit PR with clear explanation

---

📦 Packaging (for PyPI)

python -m build
twine upload dist/*

---

📄 License

MIT License © 2025 [Vinay Joshi, Avinash Raghuvanshi]

---

🧠 Philosophy

agentik isn’t trying to do everything for you — it gives you clean scaffolding and powerful abstractions to build your own intelligent agent workflows, your way.

Build smart. Stay lightweight. Go modular.

---