Multi-agent autoresearch built on git worktrees and symlinks

What it does

CORAL is infrastructure for running multiple autonomous coding agents—such as Claude Code, Codex, or OpenCode—against the same codebase in parallel. Each agent gets its own isolated git worktree, while a shared .coral/public/ directory symlinks attempts, notes, and skills into every workspace so agents see each other’s progress in real time. A grader daemon scores every commit, and a manager interrupts agents with heartbeat prompts like reflect, consolidate, or pivot to steer the collective search.

The interesting bit

Instead of building yet another agent from scratch, CORAL treats existing commercial and open-source coding agents as interchangeable workers in an evolutionary loop. The shared-state design is deliberately low-tech—symlinks and git worktrees—yet it creates a communal desk where agents spontaneously build on each other’s failed and successful commits.

Key highlights

• Native support for Claude Code, Codex, OpenCode, Cursor Agent, and Kiro; each agent runs in its own worktree with per-runtime configuration.
• Shared knowledge persists in .coral/public/ and is symlinked into every agent workspace, enabling real-time collaboration without custom network protocols.
• A grader daemon scores every commit, while a heartbeat manager interrupts agents with reflect, consolidate, or pivot prompts to steer exploration.
• Ready-to-run example tasks spanning optimization, mathematics, systems programming, and ML benchmarks (e.g., circle packing, GPU kernel optimization, Kaggle competitions).
• Backed by a research paper accepted at COLM 2026 and includes reusable LLM-judge “Rubric Judges” for open-ended tasks like reports and legal analysis.

Caveats

• Some built-in graders rely on Docker via Harbor, and the README explicitly warns that CORAL itself must not run inside Docker because Docker-in-Docker is unsupported.
• Each supported agent must be installed and authenticated separately; CORAL does not bundle the runtimes.

Verdict

Researchers and teams exploring multi-agent autoresearch or open-ended discovery should look here. If you just need a single agent to write a one-off script, the overhead of worktrees and graders is overkill.