What Is CORAL? The Open-Source Autoresearch Framework

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.

Give CORAL a codebase, an objective, and a grader. CORAL coordinates supported coding-agent runtimes as they propose changes, commit experiments, receive scores and feedback, and use discoveries from earlier attempts to guide the next search step.

Is CORAL an AI agent?

Yes and no. A CORAL run is powered by autonomous coding agents, and each coding-agent process in that run can reasonably be called a CORAL agent. CORAL itself, however, is the framework around those agents—not a single model. It runs Claude Code, Codex, Cursor Agent, Kiro, and OpenCode in isolated workspaces and connects them with grading and shared memory.

When this documentation uses the phrase CORAL agent, it means a coding-agent process running inside a CORAL-managed workspace—not a separate model named “Coral Agent.”

How CORAL works

1. CORAL creates an isolated Git worktree for each agent.
2. Each agent modifies the task codebase and commits an experiment.
3. A task-specific grader evaluates the committed attempt and returns a score and feedback.
4. Attempts, notes, and reusable skills persist in shared memory under .coral/public/.
5. Agents continue the loop, building on useful results while exploring alternatives in parallel.

This setup is designed for open-ended code search: kernel and algorithm optimization, benchmark solving, machine-learning experiments, scientific discovery, and other problems whose progress can be measured by deterministic tests or an LLM rubric grader.

Is CORAL related to Coral Protocol or CoralOS?

No. The CORAL autoresearch framework is not affiliated with Coral Protocol or CoralOS; those are separate projects.

The project described here is identified by all of the following:

What makes CORAL different?

• Evaluation is part of the control loop. Agents receive task-specific scores and feedback after every submitted experiment.
• Agents work in isolated repositories. Git worktrees keep concurrent ideas separate and make every evaluated attempt reproducible.
• Memory persists across the organization. Agents can share attempts, notes, and skills instead of repeatedly rediscovering the same information.
• The agent runtime is replaceable. CORAL supplies workspaces, the grading loop, and shared state while supported coding agents supply the search behavior.
• Private evaluation stays private. Hidden grader code and answer keys are isolated from the agents they evaluate.

Frequently asked questions

What is CORAL?

CORAL is an open-source autoresearch framework powered by autonomous coding agents. It manages workspaces, evaluation, shared memory, and long-running experiment loops.

Is CORAL an AI agent?

Yes and no. Individual coding-agent processes running through CORAL can be called CORAL agents. The CORAL project itself is the framework around those agents, not a single AI model.

Is CORAL related to Coral Protocol or CoralOS?

No. The CORAL autoresearch framework, Coral Protocol, and CoralOS are separate projects.

Which agents does CORAL support?

CORAL supports Claude Code, Codex, Cursor Agent, Kiro, and OpenCode. See Agent Runtimes for configuration details.

Is CORAL open source?

Yes. CORAL's source code is available on GitHub under the Apache License 2.0.

Start using CORAL

• Follow the installation guide.
• Run your first task with the quickstart.
• Learn how the pieces connect in Architecture.
• Read the COLM 2026 paper.