Self-Evolving Context Layers: Implicit Learning from Agent Behavior in Code Knowledge Graphs

Static code context tools serve the same responses regardless of how agents use them. We introduce gap signals — implicit metrics derived from the discrepancy between what a context tool returns and what the agent utilizes — and implement a three-loop self-improvement system in Karna, a persistent code knowledge graph for LLM agents. In a controlled A/B experiment with 70 live agent sessions (35 static, 35 adaptive, Claude Sonnet 4 via Cursor Agent CLI), the adaptive system drives 79.3% more exploratory tool calls (p=0.010) and surfaces 46.4% more entities per session (p=0.063).