MEGA: Self-Evolving AI Agent Optimization Infrastructure via Wisdom Graph

Researchers have introduced MEGA (Meta Evaluation-Grounded Adaptation), a novel infrastructure aimed at systematically improving AI agents, particularly coding agents. The core challenge addressed is moving beyond optimizing individual agents to building a system that continuously enhances them by accumulating and evolving transferable knowledge. Current approaches often lack mechanisms for compositional reasoning over accumulated knowledge or for that knowledge to self-evolve based on operational evidence. MEGA operates through a three-layered self-evolving process. Layer 1 distills reusable wisdom from agent sessions, clustering behavioral patterns and validating them empirically to create durable assets. Layer 2 then decomposes these assets into atomic PCR (Primary-Context-Resultant) units within a typed Wisdom Graph, performing deductive, abductive, and inductive reasoning to expand implicit relations and assemble context-specific execution plans. This compositional retrieval surfaces bridging knowledge that simple embedding similarity might miss. Finally, Layer 3 conducts multi-agent collaborative optimization across heterogeneous workflows, attributing improvement effects to specific strategy changes through controlled evaluation. Evidence from Layer 3 feeds back to drive the self-evolution of both the wisdom curation strategies and the optimization trajectories. This integrated approach ensures that agent system optimization and knowledge evolution are intrinsically linked, leading to more robust and intelligent agents.