CORAL

CORAL is an open-source framework from Human-Agent-Society for running organizations of autonomous coding agents that run experiments, share knowledge, and keep improving a solution. Given a codebase and a grader, it handles isolated workspaces, evaluation, and persistent shared state. The latest GitHub release listed is v0.7.21, and the project is licensed under Apache 2.0.

What It Is

CORAL is infrastructure for multi-agent autoresearch. A manager creates one Git worktree per agent plus a grader daemon. Agents read generated instructions (CORAL.md), change code, commit, and run an eval command. Each attempt is scored by a user-supplied grader, and the agent sees the score and decides its next move.

How the Loop Works

Shared state (attempts, notes, skills) lives under .coral/public/ and is symlinked into every worktree, so agents see each other's work in real time. Hidden grader state stays isolated under .coral/private/. The manager sends heartbeat prompts such as reflect, consolidate, and pivot. Version 0.6.0 added multi-island runs, which partition agents into isolated islands with scoped state and migration between islands.

Agent Runtimes and Plugin

CORAL works with Claude Code (the default), Codex, Cursor Agent, Kiro, OpenCode, Pi, and DeepSeek Harness. Each must be installed and authenticated separately. A plugin for Claude Code and Codex bundles skills that guide setup, task authoring, and running experiments from inside your own agent.

Setup Path and Examples

Install with a shell script that uses uv, then run coral init to scaffold a task and coral start -c task.yaml to launch agents. A CLI and a web dashboard (coral ui) support monitoring. Example tasks cover circle packing, a math conjecture, GPU and VLIW kernel optimization, MNIST, a Kaggle competition, and mRNA degradation prediction. Rubric-judge grader packages support open-ended tasks. The accompanying paper was accepted to COLM 2026.