GRACE: Graph-Grounded Reflective Agent Copilot Engine for Expert-in-the-Loop Knowledge Expansion

Details

Autonomous AI agents that update knowledge bases (medical ontologies, regulatory compliance databases, scientific knowledge graphs) create trust and safety problems: errors propagate silently, and experts cannot validate every agent action in real time. GRACE (Graph-Grounded Reflective Agent Copilot Engine) introduces a human-expert copilot architecture where the agent proposes knowledge graph expansions, reflects on them using graph structure to surface conflicts or gaps, and routes uncertain or high-stakes additions to expert review before committing them. The graph-grounded reflection step is the key technical contribution: the agent uses graph traversal to identify neighboring concepts that would be affected by a proposed addition, reducing the expert review burden by flagging only the consequential decisions. Evaluated on knowledge expansion benchmarks in high-stakes domains, GRACE shows improved precision on committed knowledge additions and reduced expert review load compared to flat human-in-the-loop baselines.