Source-linked AI summary
CORAL: Towards Autonomous Multi-Agent Evolution for Open-Ended Discovery
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
TL;DR
Open-ended discovery requires sustained search because optimal solutions are unknown and fixed heuristics constrain agent autonomy. CORAL enables autonomous multi-agent evolution through shared persistent memory, asynchronous agents, and heartbeat interventions, and reports stronger performance across diverse optimization tasks. The paper also identifies limitations involving frontier-model dependence, homogeneous agent initialization, and its current setting.
Problem
Open-ended problems require iterative discovery because their objectives may be clear while optimal solutions are unknown, but existing evolutionary methods fix key search and knowledge-reuse decisions.
Method
CORAL lets autonomous agents control retrieval, proposal, evaluation, and knowledge accumulation while coordinating through shared persistent memory, asynchronous execution, and heartbeat-based reflection.
Results
CORAL establishes new SOTA on 8 of 11 mathematical and systems optimization tasks, with a 2.5× higher improvement rate and 10× fewer evaluations than fixed-search baselines.
Takeaways & Limitations
The results support autonomous agents and multi-agent evolution as a promising paradigm for open-ended discovery, with multi-agent solutions exceeding single-agent results even at four times the compute.
Takeaways & Limitations
CORAL relies on frontier foundation models, initializes agents identically with shared information, and assumes the current task setting described by the authors.
Abstract
from arXiv · showhide
Large language model (LLM)-based evolution is a promising approach for open-ended discovery, where progress requires sustained search and knowledge accumulation. Existing methods still rely heavily on fixed heuristics and hard-coded exploration rules, which limit the autonomy of LLM agents. We present CORAL, the first framework for autonomous multi-agent evolution on open-ended problems. CORAL replaces rigid control with long-running agents that explore, reflect, and collaborate through shared persistent memory, asynchronous multi-agent execution, and heartbeat-based interventions. It also provides practical safeguards, including isolated workspaces, evaluator separation, resource management, and agent session and health management. Evaluated on diverse mathematical, algorithmic, and systems optimization tasks, CORAL sets new state-of-the-art results on 10 tasks, achieving 3-10 times higher improvement rates with far fewer evaluations than fixed evolutionary search baselines across tasks. On Anthropic's kernel engineering task, four co-evolving agents improve the best known score from 1363 to 1103 cycles. Mechanistic analyses further show how these gains arise from knowledge reuse and multi-agent exploration and communication. Together, these results suggest that greater agent autonomy and multi-agent evolution can substantially improve open-ended discovery. Code is available at https://github.com/Human-Agent-Society/CORAL.
1 Introduction
Open-ended discovery requires iterative search because objectives may be clear while optimal solutions remain unknown. CORAL delegates more evolutionary decisions to autonomous agents and reports stronger results across optimization tasks.
- Open-ended problems often lack ground-truth answers, so strong solutions require iterative proposal, testing, revision, and progress over time.
- Fixed evolutionary search hard-codes retrieval, testing, and knowledge-reuse decisions, motivating greater delegation of the evolutionary algorithm to autonomous agents.
- Fixed multi-agent systems commonly depend on human-defined roles and communication structures, although the optimal decomposition and interaction topology may be unknown for open-ended problems.
- CORAL uses shared persistent memory, asynchronous multi-agent organization, and heartbeat-based interventions to support autonomous long-horizon evolution.
- CORAL establishes new SOTA on 8 of 11 mathematical and systems optimization tasks, with a 2.5× higher improvement rate and 10× fewer evaluations than fixed-search baselines.
2 Related Work
Related work spans LLM-driven evolutionary search, autonomous LLM agents, and multi-agent collaboration. CORAL combines these directions for sustained, evaluator-guided optimization rather than one-off completion or stateless parallel evaluation.
- LLM-driven evolutionary-search systems embed language models in evaluator-guided loops and refine orchestration through adaptive sampling, islands, and Pareto-based selection.
- Autonomous LLM agents navigate codebases, execute code, iteratively debug, automate research cycles, and use self-feedback, tool use, or memory consolidation.
- Prior autonomous-agent systems target one-off task completion rather than sustained, goal-driven optimization.
- Multi-agent LLM systems use role assignment, structured communication, emergent cooperation, or dynamic group formation, while evolutionary systems often parallelize stateless evaluation workers without cross-step memory.
3 Coral: A Framework for Autonomous Multi-Agent Evolution
CORAL structures open-ended evolution as repeated retrieval, proposal, evaluation, and memory update, while autonomous agents control more of this process. Its shared filesystem, isolated workspaces, asynchronous organization, and heartbeat mechanisms support persistent, coordinated search.
- Problem formulation: Each improvement step retrieves context, proposes a candidate, evaluates it for score and feedback, and updates shared persistent memory.
- From fixed search to autonomous evolution: Fixed-search methods externally specify retrieval and update rules, whereas autonomous evolution lets agents decide what to retrieve, when to test or evaluate, and what to store.
- CORAL design: CORAL delegates more search decisions to agents while keeping evaluator access available through an API and addressing persistence, drift, reusable knowledge, safety, and resource challenges.
- Shared persistent memory: Shared persistent memory is implemented as a filesystem that agents access through the CORAL CLI or Bash, with progressive disclosure and folders for different knowledge types.
- Multi-agent organization: Multiple agents run asynchronously in isolated workspaces while sharing an evaluator and persistent memory, allowing parallel local exploration and cross-agent influence through stored attempts, notes, and skills.
- Heartbeat mechanisms: Heartbeat interventions support note-taking, periodic consolidation into reusable skills, and redirection after stagnation, thereby promoting explicit memory formation and reducing myopic local search.
4 Experiments
CORAL is evaluated across mathematical, systems, and stress-test optimization tasks, where it outperforms fixed evolutionary search and benefits further from multi-agent co-evolution. Analyses attribute these gains to autonomous search, knowledge accumulation, local verification, and cross-agent exploration and transfer.
- Benchmark results: CORAL achieves the best final score on all 11 mathematical and systems tasks, establishing new SOTA on 8 tasks.
- Multi-agent results: 18.3% cycle reduction on Kernel Engineering and 5.0% score increase on Polyominoes show that four-agent co-evolution extends the search frontier beyond single-agent CORAL.The gains arise from diverse search trajectories enabled by asynchronous shared persistent memory.
- Multi-agent results: 4-agent co-evolution consistently improves final scores over single-agent runs across most tasks using a fully open-source model and coding stack.The results support transfer of distributed exploration and shared persistent memory beyond proprietary models.
- Mechanistic analysis: 61% local test rate on Transaction and 57% on Kernel Engineering accompany more frequent improvement, while Prism has a 0% local test rate under randomized hidden evaluation.Local execution often catches compilation failures before external evaluation, whereas non-reproducible evaluations limit local testing.
- Mechanistic analysis: 55% of Kernel Engineering attempts improve when accessing knowledge, compared with 26% on standard tasks, where agents create only 0.05 artifacts per attempt.Advanced tasks generate 0.55 and 0.68 artifacts per attempt, including reusable bottlenecks and documented failed strategies.
- Mechanistic analysis: 36% of Kernel Engineering attempts use another agent’s commit and improve at 17% versus 9% overall, while 66% of new records originate from cross-agent parents.Polyominoes transfers more through shared notes and skills, with 87% of rounds referencing other agents’ knowledge.
- Ablations: Co-evolution outperforms the best of four independent single-agent runs on all three ablation tasks, while disabling knowledge artifacts causes an 18.6% Kernel Engineering regression.The comparison separates coordination benefits from simply adding agents or compute.
5 Conclusion
CORAL replaces rigid evolutionary search with autonomous agents that control key search and knowledge-accumulation decisions. The paper reports stronger performance from both single-agent autonomy and four-agent co-evolution.
- CORAL replaces rigid evolutionary search heuristics with autonomous agents controlling retrieval, proposal, evaluation, and knowledge accumulation.
- A single autonomous agent outperforms fixed evolutionary-search baselines across the evaluated task suites.
- Four-agent co-evolution discovers solutions that no single agent finds, even when the single agent receives four times the compute.
- The results suggest autonomous agents are a promising paradigm for open-ended discovery requiring iterative search, feedback, and knowledge accumulation.
LLM Usage Disclosure
The authors used LLMs only for minor writing assistance, while CORAL agents used LLMs as the backbone of the experiments. The authors conducted the research and produced its substantive outputs.
- LLMs provided minor writing assistance, including grammar correction and language polishing.
- CORAL agents were instantiated with LLMs as the backbone for agent behaviors and experiments.
- The authors developed and carried out the research ideas, methodology, experiments, implementation, analysis, and conclusions.
- No LLM generated the research ideas, experimental results, figures, or evaluations.
A Limitations and Future Directions
CORAL’s current limitations concern model requirements, agent homogeneity, and reliance on reasonably well-specified evaluators. Future work targets smaller specialized models, heterogeneous agents, and co-evolving evaluation.
- CORAL relies on frontier foundation models capable of complex coding-agent workflows, making full local deployment difficult.
- All agents currently share identical initialization and information, so future work could introduce distinct roles, personalities, or private information.
- The current setting assumes a reasonably well-specified evaluator, whereas some open-ended problems have incomplete, difficult, or ambiguous evaluators.
- Future evaluation could co-evolve with solutions through evaluator refinement, learned critics, or human-agent negotiation.
B Additional Experiment Results
The Polyominoes packing experiment evaluates how tightly agents can pack all polyominoes into a grid. CORAL with four agents substantially exceeds both a single-attempt baseline and the previous state of the art.
- The Polyominoes packing problem packs all polyominoes tightly into a grid to minimize unused area.
- 89.4% coverage is achieved by four-agent CORAL with Claude Opus 4.6 and web search, surpassing the previous SOTA of 87%.
- 56.0% coverage is achieved by a single-attempt Claude Opus 4.6 baseline.
B.2 Trajectory Statistics for Autonomous Self-Evolution
This section defines trajectory statistics for analyzing autonomous self-evolution and reports how local verification and knowledge reuse relate to improvement across tasks.
- Trajectory analysis measures improvement, local testing, prior-work inspection, knowledge creation, and knowledge access during search.Impr. Rate measures attempts improving over the current best; Test→Impr. and Read→Impr. measure improvement among locally tested or knowledge-accessing attempts.
- 89.4% is the CORAL score for Polyominoes packing, versus 56.0% for the single-attempt baseline and 87% for the previous best known score.The figure uses Claude Opus 4.6 through Claude Code with web search access.
- Local verification and knowledge reuse are strongly associated with successful improvement, but their frequency varies substantially across tasks.Tasks with cheap, reliable local testing benefit more from verification, while tasks with richer reusable insights show higher knowledge creation and access rates.
- Table 4 summarizes trajectory statistics for standard tasks on average and for two advanced benchmarks.
C Additional Implementation Details
CORAL’s implementation combines structured agent prompts, shared-memory workflows, heartbeat interventions, and a command-line interface for autonomous iterative search.
- C.1 Prompts: Each agent receives a generated CORAL.md containing task-specific instructions and system interface documentation.The file is instantiated from single-agent or multi-agent templates using task name, description, score direction, shared directory, and agent ID.
- C.1 Prompts: Agents follow a plan–edit–evaluate–repeat workflow while inspecting prior attempts, notes, skills, and leaderboard results.After evaluation, agents update shared notes and skills to record findings and reusable techniques.
- Heartbeat mechanisms: Heartbeat prompts trigger reflection every evaluation, consolidation every 10 evaluations, and pivots after five non-improving evaluations.Pivot interventions diagnose plateaus, inspect top attempts, choose a different direction, and restart from a strong base.
- Heartbeat mechanisms: Consolidation heartbeats require synthesis notes, a connections map, and an open-questions list in shared memory.The process asks agents to synthesize findings, connect patterns, and document contradictions or gaps.
- System interface: CORAL exposes agent functionality through a CLI with 17 commands organized into four categories, including the evaluation pipeline.Agents interact with CORAL exclusively through these commands.
D.2 Task Configurations
CORAL evaluates diverse mathematical, systems, and stress-test tasks while correcting evaluator defects that could distort benchmark scores.
- Configurations: Representative configurations illustrate the range of mathematical and systems optimization tasks supported by CORAL.
- Task diversity: The evaluation spans six mathematical optimization tasks, five systems optimization tasks, and Kernel Engineering and Polyominoes stress tests.Grading uses subprocess execution, constraint validation, benchmark-relative scoring, and external evaluation frameworks.
- Evaluator corrections: Table 8 summarizes evaluator bug fixes applied to the ADRS benchmark evaluators to ensure correct scoring.
- Evaluator corrections: Evaluator corrections ensure that failed, invalid, dropped, or type-incompatible cases receive appropriate handling rather than artificially favorable scores.Fixes address PRISM failures, invalid transaction schedules, dropped EPLB experts, and mixed-type LLM-SQL inputs.
E.1 Setup Details
Experiments compare CORAL with fixed evolutionary search baselines under matched models, evaluators, and time budgets across mathematical, systems, and stress-test tasks.
- Hardware and models: All experiments ran on Linux machines, using CPU-only instances because the evaluated tasks rely on CPU execution or simulation.
- Hardware and models: Claude Opus 4.6 is the primary backbone for CORAL and all baselines, while MiniMax M2.5 supports open-source generalization experiments.Internet access is enabled only when a task configuration explicitly sets the research flag.
- Baselines: The baselines are OpenEvolve, ShinkaEvolve, and EvoX, all receiving identical seed programs, evaluators, and wall-clock budgets.
- Evaluation protocol: Mathematical and systems suites use 3-hour budgets averaged over four runs, while stress tests stop after 100 non-improving evaluations or two hours.Multi-agent experiments use four agents with matched wall-clock time.
- Efficiency and cost: Four-agent runs cost approximately 3–4× a single-agent run, while typical three-hour single-agent mathematical runs cost about $30–60.
- Efficiency and cost: CORAL’s improvement rate is substantially higher than structured baselines while typically using fewer evaluation calls within the same wall-clock budget.The infrastructure adds negligible overhead; dominant costs are LLM API calls and grader execution.