GenericAgent: Maximizing Context Information Density for Self-Evolving AI Agents

I will explain **GenericAgent(GA)**, a self-evolving AI agent system that maximizes the density of information by each situation and increases efficiency. In order to overcome the limitations of existing systems that reduce the quality of reasoning due to vast contexts, GA adopts a minimal set of tools and a hierarchical on-demand memory structure to selectively utilize only essential information. In particular, self-evolving mechanisms that compress verified work paths into standard operating procedures (SOPs) and workable code dramatically reduce token usage and improve performance in repetitive tasks. It also applies context cutting and compression technology to maintain high information density and ensure high success rates even when performing long-term tasks. As a result of the experiment, GA demonstrated excellent work completion ability at a much lower cost than conventional agents on major benchmarks. In conclusion, the study presents a new design paradigm, emphasizing that the ability of the agent depends on the density of decision-related information, not the length of the context. https://arxiv.org/pdf/2604.17091