# Self‑correcting LangGraph Agent

This repository contains a small demo of a **self‑correcting LangGraph agent**. The agent receives a *task* string, executes it via an unreliable tool, then asks an LLM to judge the result. If the judge says the result is **failed**, the agent retries until it reaches a maximum number of attempts.

## Features

- **Unreliable tool** – 30 % chance of raising an exception.
- **LLM judge** – forces the model to answer only `success` or `failed`.
- **Retry logic** – automatically retries until success or a maximum number of attempts.
- **LangGraph** – low‑level graph with three nodes: `execute_task`, `verify_result`, `handle_error`.

## Setup

```bash
# Optional: create a virtual environment
python -m venv venv
source venv/bin/activate   # Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt
```

## Running the Agent

```bash
python agent.py
```

You will be prompted to enter a task. The agent will then perform the task, verify the result, and retry if necessary. Example output:

```
Self‑correcting LangGraph agent demo
Enter a task: 2+2

--- Result ---
Task: 2+2
Attempts: 2
Status: success
Result: 22
```

## Project Structure

- `agent.py` – main implementation.
- `README.md` – this documentation.
- `requirements.txt` – Python dependencies.

## Notes

- The LLM used is OpenAI's `gpt-4o-mini`. If you prefer Ollama, change the `ChatOpenAI` import to `ChatOllama` and adjust the model name accordingly.
- The unreliable tool is a toy example; replace it with a real tool for production use.