[2604.01658] CORAL: Towards Autonomous Multi-Agent Evolution for Open-Ended Discovery

CORAL (arXiv:2604.01658) is a framework for autonomous multi-agent evolution that replaces rigid, hard-coded heuristics with long-running LLM agents that explore, reflect, and collaborate through shared persistent memory, asynchronous execution, and heartbeat-based interventions.

By delegating search decisions to agents that accumulate reusable notes and skills, CORAL achieves state-of-the-art results on 10 out of 11 mathematical and systems optimization tasks, with 3–10× higher improvement rates and faster convergence than fixed evolutionary baselines. Notably, on Anthropic’s kernel engineering task, four co-evolving agents improved the best known score from 1363 to 1103 cycles.

Key architectural components include: Shared Persistent Memory: A common repository where agents store attempts, notes, and skills, enabling knowledge reuse and indirect collaboration. Asynchronous Multi-Agent Organization: Agents operate in isolated workspaces while coordinating through the shared memory, allowing for diverse exploration trajectories. * Heartbeat Mechanisms: Periodic interventions that prompt agents to reflect on progress, consolidate knowledge, and redirect search efforts.

CORAL is presented as an architecture for autonomous multi-agent evolution in the context of open-ended scientific discovery. Rather than treating discovery as a single-agent, single-objective optimization problem, CORAL frames it as a population-based process in which multiple agents generate, modify, evaluate, and combine hypotheses, experiments, or research strategies over time. The system is designed to maintain a persistent record of candidate ideas and outcomes, allowing later agents to build on earlier results, avoid redundant search, and preserve diverse lines of inquiry.

A central contribution is the integration of multi-agent reasoning with evolutionary search mechanisms such as variation, selection, recombination, and novelty preservation. In this model, agents can propose new research directions, critique or refine existing ones, and adapt their behavior based on accumulated evidence, while evolutionary pressure helps balance exploration with exploitation. The emphasis on open-endedness suggests that CORAL is not merely optimizing a fixed benchmark or narrow reward function, but is intended to support sustained discovery in problem spaces where the space of valuable outcomes is not fully known in advance.

This matters because many important scientific and technical discovery tasks are ill-defined, long-horizon, and resistant to purely goal-directed optimization. By coupling modern AI agents with evolutionary dynamics, CORAL offers a potential path toward more cumulative, diverse, and self-sustaining automated research systems. More broadly, the work is significant as a design pattern for autonomous discovery pipelines: one in which progress emerges not from a single model solving a predefined task, but from the interaction of many agents operating within a shared evolutionary ecosystem.