CORAL: Towards Autonomous Multi-Agent Evolution for Open-Ended Discovery by Ao Qu, Han Zheng, Zijian Zhou, Yihao Yan, Yihong Tang, Shao Yong Ong, Fenglu Hong, Kaichen Zhou, Chonghe Jiang, Minwei Kong, Jiacheng Zhu, Xuan Jiang, Sirui Li, Cathy Wu, Bryan Kian Hsiang Low, Jinhua Zhao, Paul Pu Liang (2026)

tags

Open-ended evolution, Co-evolution, Quality diversity, Genetic algorithms, Coding agent, LLM

source

(Qu et al. 2026)

Summary

CORAL is a framework for autonomous multi-agent evolution applied to open-ended discovery problems. Unlike prior LLM-driven evolutionary search systems (FunSearch, AlphaEvolve, OpenEvolve) that rely on fixed heuristics for parent selection, mutation, and population management, CORAL delegates these decisions to autonomous LLM agents. Each agent operates in an isolated workspace, iteratively proposing, evaluating, and refining candidate solutions while reading from and writing to a shared persistent memory system.

The framework introduces three core mechanisms: (1) shared persistent memory structured as a file system containing attempts, notes, and reusable skills; (2) asynchronous multi-agent organization where agents coordinate indirectly through the shared memory rather than via explicit messaging; and (3) heartbeat-based interventions that periodically prompt agents to reflect, consolidate knowledge, and redirect when stuck. This design moves from fixed evolutionary search (where the LLM is just a mutation operator) through autonomous single-agent evolution to autonomous multi-agent co-evolution.

Evaluated on 11 mathematical, algorithmic, and systems optimization tasks, CORAL achieves the best final score on all tasks, setting new SOTA on 8. Its improvement rate is 3-10x higher than fixed evolutionary baselines while using far fewer evaluations (5-20 vs. 60-100). On Anthropic’s kernel engineering stress test, four co-evolving agents reduce execution cycles from 1,363 to 1,103 (a 20% gain). Ablation studies confirm that both knowledge accumulation and multi-agent co-evolution contribute causally to performance—co-evolution outperforms the best-of-4 independent single-agent runs.

Key Ideas

• Distinguishes three paradigms for LLM-based open-ended discovery: fixed evolutionary search, autonomous single-agent evolution, and autonomous multi-agent evolution, with increasing agent autonomy at each level.
• Shared persistent memory as file system: agents store scored attempts, free-form notes (observations, failure modes), and reusable skills (scripts + natural-language descriptions) in a shared directory, enabling knowledge diffusion without explicit communication protocols.
• Heartbeat mechanism with three types: per-iteration reflection (write notes), periodic consolidation (merge findings into skills), and stagnation-triggered redirection (pivot strategy when stuck). These replace the fixed search loop of prior methods.
• Agents exhibit emergent collaborative behaviors: cross-agent information transfer accounts for 36% of attempts on kernel engineering, with 66% of new records originating from cross-agent parents.
• Exploration diversity is high: pairwise Jaccard similarity of strategy keywords is only 0.43 on kernel engineering and 0.31 on polyominoes, meaning each agent explores substantially different regions of the solution space.
• Knowledge accumulation is causally important: disabling notes and skills causes an 18.6% regression on kernel engineering (1,350 to 1,601 cycles).
• All agents use Claude Opus 4.6 via Claude Code, and multi-agent results also generalize to open-source models (MiniMax M2.5 + OpenCode).

Comments

CORAL represents a natural evolution of the LLM-as-mutation-operator paradigm established by FunSearch and AlphaEvolve, pushing toward full agent autonomy in the evolutionary loop. The key insight is that for complex open-ended problems, the search decisions themselves (what to explore, what to remember, when to pivot) are as important as the mutations, and LLM agents are capable of making these decisions autonomously.

The shared persistent memory design is reminiscent of quality diversity archives like MAP-Elites, but unstructured and agent-curated rather than algorithmic. The heartbeat mechanism for reflection and redirection parallels novelty search ideas about escaping local optima, but implemented through natural-language self-reflection rather than behavioral distance metrics.

The connection to open-ended evolution is direct: CORAL is attempting to create a system where the search process itself evolves and adapts, with agents accumulating knowledge and skills over time. This relates to the AI-generating algorithm vision of meta-learning systems that improve their own learning process.

A notable limitation is the reliance on frontier foundation models—the framework requires agents capable of complex coding-agent workflows. The authors note that all agents are initialized identically, leaving room for bootstrapped heterogeneity through distinct personalities or private information.

Connections

• Related to Open-ended evolution because CORAL directly targets open-ended discovery problems requiring sustained search and unbounded improvement.
• Related to Co-evolution because multiple agents co-evolve solutions asynchronously, sharing discoveries through persistent memory.
• Related to Quality diversity and MAP-Elites because the shared memory acts as an unstructured quality-diversity archive, and agents maintain exploration diversity.
• Related to Coding agent because CORAL agents are LLM coding agents that write, test, and iterate on code solutions.
• Related to AI-GAs: AI-generating algorithms because CORAL moves toward systems that autonomously improve their own search process.
• Related to POET because both use co-evolutionary dynamics for open-ended improvement, though CORAL co-evolves solutions via shared memory rather than co-evolving environments and agents.
• Related to Meta-Harness because both use LLM coding agents to iteratively optimize programs through evolutionary search.
• Related to Novelty search because the heartbeat redirection mechanism serves a similar purpose to novelty-driven exploration—escaping local optima.
• Related to Program synthesis because the tasks involve evolving programs (kernels, heuristics, algorithms) through LLM-guided search.

Bibliography

1. Ao Qu, Han Zheng, Zijian Zhou, Yihao Yan, Yihong Tang, Shao Yong Ong, Fenglu Hong, et al.. April 2, 2026. "CORAL: Towards Autonomous Multi-Agent Evolution for Open-Ended Discovery". https://arxiv.org/abs/2604.01658.