Autonomous multi-branch research lab. Branches compete for compute budget, converging on what works. The system is a local-first runtime for AI agents like Claude Code and Codex, tackling real research problems with structure for long runs. This system is designed for researchers and engineers seeking to automate and optimize AI/ML exploration by rewarding reproducible progress and robust evaluation.

How It Works

The core mechanism employs asynchronous population search, where multiple research "families" or idea branches compete. Compute is allocated via a "funding" system, minting credits for stable, reproducible progress and spending them on new descendants. Evaluation is consistently performed externally, with "decisive challenges" (held-out tests) granting families strategic importance beyond local metric optimization. A supervisor-worker model manages tasks, utilizing a "File-as-Bus" workspace for durable state persistence via files and append-only logs, allowing thin control over thick project state.

Quick Start & Requirements

• Primary install: pip install -e '.[nlp-sentiment]' within a Python virtual environment.
• Prerequisites: Python 3, virtual environment. Requires Claude Code or Codex for agent operation.
• Example Setup: The examples/nlp-sentiment/research_lab provides a runnable dashboard and full scaffold in approximately 5 minutes, including labrat doctor, labrat bootstrap, and serving a local HTTP server.
• Relevant Links: program.md, docs/DEEP_RESEARCH.md.

Highlighted Details

• Population Search & Funding: Families of ideas compete; compute is awarded based on reproducible progress and winning decisive challenges.
• Decisive Challenges: Families gain status by winning held-out tests, ensuring robustness beyond local metric overfitting.
• Unified Agent Interface: Claude Code and Codex are treated as peer operators with a consistent runtime contract and file layout.
• File-as-Bus Workspace: State is managed durably through files and append-only logs, simplifying supervisor control.
• Operator Surfaces: Provides structured interfaces (Markdown files, slash commands for Claude Code) for agent interaction.

Maintenance & Community

The project originated from DXRG. Further profiles are planned for future releases. No specific community channels (e.g., Discord, Slack) or detailed roadmap links are provided in the README.

Licensing & Compatibility

The license type is not specified in the provided README text, which is a significant omission for due diligence.

Limitations & Caveats

The system is best suited for problems with a clear baseline, a bounded experiment runner, a consistently measurable metric, and at least one hard held-out challenge. It is framed as a tool for structured research exploration rather than a general philosophy-of-science engine. Ongoing development is indicated by planned future profile additions.