Harnessing Agentic Evolution

arxiv:2605.13821

Original source · Paper JSON

Agents, open-endedness, AGI · 2026-05-13

Cites (8) · Cited by (0)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

1. Part 1
2. Part 2
3. Part 3

Contents

• Harnessing Agentic Evolution
• Abstract
• 1 Introduction
• 2 Related Work
• 3 Problem Formulation
• 3.1 Agentic Evolution
• 3.2 Evolution as an Interactive Environment
• 4 Methodology
• 4.1 Design of AEvo
• 4.2 Instantiating AEvo
• 5 Experiments
• 5.1 Settings
• 5.2 Main Results
• 5.3 Evolution Dynamics
• 5.4 Ablation Study
• 5.5 Case Study
• 6 Conclusion
• Appendix A Ablation Study Details
• Appendix B Implementation Details
• Appendix C Additional Experimental Details
• C.1 Task and Metric Details
• Circle Packing.
• Autocorrelation-Second.
• Performance Engineering.
• Reporting Protocol.
• Cost Computation.
• C.2 Initialization Details
• C.2.1 Initial Procedure in Procedure-Based AEvo
• C.2.2 Initial Agent in Agentic and Reasoning Tasks
• C.2.3 Meta-Agent Skill Excerpt
• C.3 Representative Evolved Artifacts and Outcomes
• C.3.1 Representative Evolved Procedures
• C.3.2 Representative Evolved Agent Harness
• Task skill.
• Session goal.
• Persistent family map.
• Replay utility.
• Session memory written back by the inner agent.
• C.3.3 Representative Optimization Outcomes
• ARC-AGI-2 best artifact.
• Analysis.
• Performance-engineering best artifact.
• Submitted wrapper.
• Parameterized benchmark family.
• Core benchmark schedule.
• Analysis.
• Instructions for reporting errors