CORAL Documentation

An orchestration system for autonomous coding agents. Agents follow a CORAL.md guide, run experiments, share knowledge, and loop forever.

Core Pattern

Spawn agents → agents read CORAL.md → commit changes → eval runs → repeat

CORAL manages the full lifecycle: creating isolated workspaces, generating agent instructions, running evaluations, and collecting results.

How It Works

coral start --config task.yaml
  → Creates .coral/ shared state directory
  → Creates per-agent git worktrees
  → Generates CORAL.md in each worktree
  → Spawns coding agents (Claude Code, Codex, OpenCode)

Each agent:
  → Reads CORAL.md for instructions
  → Makes changes, commits
  → Runs `coral eval -m "description"`
  → Eval writes attempt JSON to .coral/attempts/
  → Agent sees score, decides next move
  → Shares notes in .coral/notes/
  → Packages tools as skills in .coral/skills/

Key Concepts

• Agents are the optimizers — Claude Code, Codex, or OpenCode subprocesses working in git worktrees
• Shared state via .coral/ — attempts, notes, skills symlinked into each worktree
• Eval loop — agents call coral eval -m "..." to stage, commit, and grade
• CLI orchestration — coral start/stop/status/eval/log and more
• Web dashboard — coral ui for real-time monitoring

Quick Links

• Installation — Get CORAL running
• Quick Start — Your first task in 5 minutes
• Configuration — Full config reference
• CLI Reference — All commands
• Writing Graders — Build your own evaluation