CORAL provides a robust, lightweight infrastructure for multi-agent autonomous systems, designed for autoresearch and self-evolution. It enables AI agent organizations to collaboratively run experiments, share knowledge, and iteratively improve solutions. By automating isolated workspaces, safe evaluations, and persistent knowledge sharing, CORAL allows researchers and power users to build self-improving AI with reduced configuration overhead, facilitating open-ended discovery.

How It Works

Agents operate within isolated Git worktrees, sharing knowledge (attempts, notes, skills) in real-time via symlinked directories from a central .coral/public/, eliminating sync overhead. A manager orchestrates agents, capable of triggering reflective prompts. Agents use uv run coral eval for staging, committing, and grading work. Orchestration is managed via an extensive CLI and a real-time web dashboard. An optional "warm-start" feature allows agents to perform initial web-based literature reviews before coding.

Quick Start & Requirements

• Installation: Clone repo, cd CORAL, uv sync.
• Requires Python 3.11+.
• Critical Prerequisite: Users must independently install, configure, and authenticate AI coding agents (e.g., Claude Code, Codex, OpenCode). CORAL does not manage agent setup.
• Launch: uv run coral start -c <task_config.yaml>.
• Web dashboard: uv sync --extra ui && uv run coral ui.
• Official examples and documentation are included.

Highlighted Details

• Native integration with Claude Code, Codex, and OpenCode.
• Facilitates autonomous self-evolution and autoresearch via multi-agent collaboration.
• Real-time shared knowledge through symlinked directories.
• Web dashboard for live monitoring and agent status.
• Optional "warm-start" for literature review.
• Built-in LiteLLM gateway for custom model routing and API management.

Maintenance & Community

• Associated with a recent arXiv paper (2604.01658v1) and blog post.
• Development appears recent.
• Supported by TNT Accelerator.
• No explicit community channels (Discord/Slack) or public roadmap detailed in the README.

Licensing & Compatibility

• MIT License.
• Permissive for commercial use, modification, and distribution.

Limitations & Caveats

• User is solely responsible for agent installation, configuration, and authentication; system failure occurs if agents are not properly set up.
• Operational costs (budget) for multiple agents should be considered.
• Effectiveness depends on the quality of the provided codebase, grading script, and integrated AI agents.