This project provides an autonomous AI agent designed to automate the execution and monitoring of deep learning experiments, enabling researchers to focus on higher-level thinking. It targets researchers who understand their experimental goals but wish to offload the repetitive tasks of running, observing, and iterating on experiments, thereby maximizing research time and enabling continuous, 24/7 operation.

How It Works

The core of the system operates on a THINK → EXECUTE → REFLECT loop. The agent analyzes a PROJECT_BRIEF.md file, plans the next experimental step, modifies code and configurations, executes a dry-run, and then launches the training job. During training, it employs a novel zero-cost monitoring strategy, using only process checks and log reads without invoking LLMs, thus drastically reducing operational expenses. Once training concludes, the agent enters the REFLECT phase, parsing results, comparing them against baselines, and deciding on the subsequent action. This process repeats autonomously. A Leader-Worker architecture manages tasks, and a two-tier memory system (PROJECT_BRIEF.md and MEMORY_LOG.md) ensures a constant, manageable context size.

Quick Start & Requirements

Installation involves cloning the repository, installing Python dependencies via pip install -r requirements.txt, and then installing Claude Code slash commands using python install.py. Key requirements include Python 3.10+, at least one NVIDIA GPU, and an API key for Anthropic or OpenAI. A PROJECT_BRIEF.md file is mandatory for defining experiments, with an optional config.yaml for overriding defaults. Obsidian sync is supported for enhanced progress tracking.

Highlighted Details

• Zero-Cost Monitoring: Achieves significant cost savings by performing LLM-free monitoring during training, with LLM calls only during the THINK and REFLECT phases, resulting in approximately $0.08 per 24-hour cycle.
• Battle-Tested Operation: Demonstrates robustness with over 500 autonomous experiment cycles completed, significant single-project improvements (52% over baseline), and continuous operation for over 30 days without human intervention.
• Constant-Size Memory: Utilizes a fixed-size, two-tier memory structure to manage project context and results, preventing unbounded growth over long-running experiments.
• Human-in-the-Loop Control: Maintains researcher oversight through a PROJECT_BRIEF.md for stable goals and constraints, and HUMAN_DIRECTIVE.md or command-line flags for temporary steering.

Maintenance & Community

Recent updates in April 2026 indicate active development and maintenance. While specific community links like Discord or Slack are not detailed, the project is hosted on GitHub, facilitating contributions and issue tracking.

Licensing & Compatibility

The project is released under the Apache 2.0 license, which is permissive and generally compatible with commercial use and linking in closed-source projects.

Limitations & Caveats

This agent is designed as an operator for experiment execution, not a replacement for scientific judgment, interpretation, or hypothesis generation. The core philosophy emphasizes keeping the human researcher in control of the research direction and interpretation, with the agent handling the repetitive operational tasks.