README.md

---
title: MAKER Agent
emoji: 🔧
colorFrom: indigo
colorTo: purple
sdk: gradio
app_file: app.py
pinned: false
license: mit
short_description: Reliable AI Agent with Voting
---

# 🔧 MAKER Agent

**Reliable AI with Voting & Red-Flagging**

Based on the paper: [Solving a Million-Step LLM Task with Zero Errors](https://arxiv.org/abs/2511.09030)

## What is MAKER?

MAKER (Massively Decomposed Agentic Processes) is a framework that achieves **near-zero errors** over long task sequences by:

| Pillar | Description |
|--------|-------------|
| **Maximal Decomposition** | Break tasks into single-step atomic operations |
| **K-Voting** | Sample multiple times, winner needs K votes ahead |
| **Red-Flagging** | Discard suspicious outputs (don't repair them) |

## Key Insight

> *"Reliability is an engineering problem, not a model problem."*

Instead of waiting for better models, you can achieve near-zero errors TODAY using smaller, cheaper models with statistical voting.

The paper achieved **1,000,000 steps with ZERO errors** using gpt-4.1-mini!

## Features

- 🔍 **Web Search**: Optional DuckDuckGo search (free, no API key)
- 📁 **File Upload**: PDF, DOCX, TXT, images, and more
- 🗳️ **Voting**: Configurable K parameter for reliability vs speed
- 🚩 **Red-Flagging**: Automatic detection and discarding of suspicious outputs
- 🔌 **Multiple LLMs**: OpenAI, Anthropic, Groq, Together, OpenRouter

## Usage

1. **Setup**: Enter your API key and configure K (votes needed)
2. **Query**: Ask questions with optional web search or file context
3. **Results**: See the winning answer with vote distribution

## Supported LLM Providers

| Provider | Models |
|----------|--------|
| OpenAI | gpt-4o-mini, gpt-4o, etc. |
| Anthropic | claude-sonnet, claude-opus |
| Groq | llama-3.3-70b (fast!) |
| Together | Llama, Mistral, etc. |
| OpenRouter | 100+ models |

## Local Development

```bash
git clone https://huggingface.co/spaces/YOUR_USERNAME/maker-agent
cd maker-agent
pip install -r requirements.txt
python app.py
```

## Paper Citation

```bibtex
@article{meyerson2025maker,
  title={Solving a Million-Step LLM Task with Zero Errors},
  author={Meyerson, Elliot and others},
  journal={arXiv preprint arXiv:2511.09030},
  year={2025}
}
```

## License

MIT