Isolated workspaces
Run agents in independent git worktrees so every idea can evolve without disrupting the others.
Open infrastructure for autoresearch
Open-source autoresearch powered by autonomous coding agents.
CORAL is an open-source autoresearch framework powered by autonomous coding agents. It runs long-lived agents in isolated Git worktrees, continuously grades experiments, and shares knowledge through persistent memory.
The CORAL loop
spawn agents
↓
commit experiments
↓
grade every attempt
↓
share what works
↺Built for continuous improvement
The infrastructure around the agents
Continuous grading
Score every committed attempt with your own grader and feed concrete results back into the loop.
Shared knowledge
Carry attempts, notes, and reusable skills across agents through a persistent shared state.
How CORAL works
A filesystem-native evolution loop
The manager creates one worktree per agent and a grader daemon for evaluation. Agents submit commits, receive scores, and exchange discoveries through .coral/public/. Hidden grader state stays isolated under .coral/private/.
- 1Create isolated agent worktrees
- 2Commit and submit an experiment
- 3Grade the attempt safely
- 4Share results and continue
Quick Start
Start an agent organization
Install CORAL, scaffold a task, and launch your first run from a small YAML configuration.
Read the Quick Startcurl -fsSL https://raw.githubusercontent.com/Human-Agent-Society/CORAL/main/install.sh | sh
coral init my-task
cd my-task
coral start -c task.yamlSupported agents
Bring your coding agent
Claude CodeCodexCursor AgentKiroOpenCode
Configure agent runtimes →Examples
Explore across domains
OptimizationMathematicsSystemsGPU kernelsMLBio/ML
Browse example tasks →Research
CORAL at COLM 2026
Read the paper on autonomous multi-agent evolution for open-ended discovery.
Read the paper →Latest blog
Evolve Like Coral: Towards Autonomous Multi-Agent Evolution
CORAL experiments on autonomous multi-agent evolution, collaboration, and open-ended discovery.
Read the article →