CORAL: MIT Built a Colony of Self-Evolving Agents for Autoresearch
After three weeks of tracking this paper on HuggingFace, the code is finally out. And it's worth the wait.
CORAL is a multi-agent framework from MIT, Stanford, NUS, and Meta Superintelligence Lab. The premise is simple but powerful: instead of one agent iterating on a problem, you spin up multiple agents that explore independently, share knowledge through persistent memory, and evolve their approaches through collaboration. Give it a codebase and a grading script, and CORAL handles the rest.
The architecture is clever. Each agent runs in its own git worktree branch, so there's no interference between experiments. Shared state is maintained through symlinked directories, so agents can see each other's work in real time without synchronization overhead. A heartbeat system monitors agent health and intervenes when things go off track. It supports Claude Code, OpenCode, and Codex as backend agents.
The results justify the design. On Anthropic's kernel engineering task, four co-evolving agents improved the best known score from 1363 to 1103 cycles. CORAL's improvement rate is 3-10x higher than fixed evolutionary search, and it typically converges within 5-20 evaluations versus 60-100 for baseline methods. That's not incremental. That's a different class of performance.
What makes CORAL different from prior multi-agent systems is the emphasis on genuine autonomy. Agents don't just run tasks in parallel, they decide what to try, reflect on what worked, and teach each other through a shared skill and notes system. It's closer to a research lab than a task queue.
The project includes a web dashboard for monitoring, 17+ CLI commands for management, and built-in LiteLLM proxy support. MIT licensed. Code at github.com/Human-Agent-Society/CORAL. Paper at arxiv.org/abs/2604.01658.
← Back to all articles
CORAL is a multi-agent framework from MIT, Stanford, NUS, and Meta Superintelligence Lab. The premise is simple but powerful: instead of one agent iterating on a problem, you spin up multiple agents that explore independently, share knowledge through persistent memory, and evolve their approaches through collaboration. Give it a codebase and a grading script, and CORAL handles the rest.
The architecture is clever. Each agent runs in its own git worktree branch, so there's no interference between experiments. Shared state is maintained through symlinked directories, so agents can see each other's work in real time without synchronization overhead. A heartbeat system monitors agent health and intervenes when things go off track. It supports Claude Code, OpenCode, and Codex as backend agents.
The results justify the design. On Anthropic's kernel engineering task, four co-evolving agents improved the best known score from 1363 to 1103 cycles. CORAL's improvement rate is 3-10x higher than fixed evolutionary search, and it typically converges within 5-20 evaluations versus 60-100 for baseline methods. That's not incremental. That's a different class of performance.
What makes CORAL different from prior multi-agent systems is the emphasis on genuine autonomy. Agents don't just run tasks in parallel, they decide what to try, reflect on what worked, and teach each other through a shared skill and notes system. It's closer to a research lab than a task queue.
The project includes a web dashboard for monitoring, 17+ CLI commands for management, and built-in LiteLLM proxy support. MIT licensed. Code at github.com/Human-Agent-Society/CORAL. Paper at arxiv.org/abs/2604.01658.
Comments