CORAL Documentation

CORAL is an open infrastructure designed for organizing and managing autonomous agents that experiment, learn, and evolve collectively. It addresses the challenge of enabling multi-agent organizations to continuously improve solutions to open-ended problems by providing a structured environment where agents can operate independently yet share knowledge and results.

The platform offers several core features to support this process. Each agent operates within an isolated workspace, implemented as an independent git worktree, allowing for the parallel evolution of ideas without interference. CORAL includes a continuous grading mechanism, where every committed experiment by an agent is evaluated using a customizable grader daemon. This setup ensures that concrete results are fed back into the improvement loop, promoting ongoing optimization and learning. Shared memory is another key capability, enabling agents to exchange attempts, notes, and reusable skills through a persistent shared state. Discoveries and results are made available to all agents via a public directory, while sensitive grader information remains protected in a private directory.

The workflow in CORAL involves spawning agents, committing experiments, grading each attempt, and sharing what works across the organization. The infrastructure is filesystem-native, emphasizing the use of the local file system and git for managing agent worktrees and experiment tracking. Setting up an agent organization in CORAL involves installing the platform, scaffolding a task, and configuring the initial run through a YAML file.

CORAL is intended for scenarios where autonomous agents must collaborate, experiment, and evolve solutions in a continuous, reproducible manner.