Human-Agent-Society/CORAL

What it solves

CORAL provides a robust infrastructure for autonomous AI agent organizations to conduct autoresearch. It allows multiple coding agents to run experiments, share knowledge, and continuously improve solutions to a given task—such as optimizing code performance or solving mathematical conjectures—without manual intervention.

How it works

CORAL manages a fleet of AI coding agents (e.g., Claude Code, Cursor Agent, Codex) each operating in its own isolated git worktree. These agents share a common state directory (.coral/public/) where they can exchange notes, skills, and previous attempts. A dedicated grader daemon continuously scores every commit, providing a feedback loop that guides the agents toward the optimal solution. A manager interrupts agents with periodic prompts to reflect, consolidate, and pivot their strategies.

Who it’s for

It is designed for researchers and developers who want to automate the iterative process of discovery and optimization through multi-agent collaboration and self-evolution.

Highlights

• Multi-Agent Collaboration: Supports a wide range of agent runtimes including Claude Code, OpenCode, and DeepSeek Harness.
• Symmetric Shared State: Agents see each other's work in real-time via symlinked public directories.
• Safe Evaluation: Integrates with Docker for isolated, unprivileged user execution to prevent agents from accessing private grader keys.
• Extensible Graders: Supports custom grader packages and LLM-judge rubric judges for open-ended tasks.
• Plugin Integration: Offers plugins for Claude Code and Codex to scaffold and launch tasks directly from within the agent's interface.