Data agents are designed to carry out natural-language instructions across heterogeneous sources such as tables, files, and databases. Yet they face what the authors call an agent-data gap: the data lives outside the agent, and the agent can only reach it through generic tools that expose limited signals like column names or file paths.
Existing remedies tend to fall into two camps. Some let agents explore raw data sources directly; others inject manually constructed semantic layers into prompts. According to the paper, neither approach scales well to large heterogeneous sources, and neither adapts to differing agent behaviors.
EvoOntology is presented as a self-evolving ontology layer for data agents. It encapsulates the ontology as an MCP server made up of a schema layer, a content layer, and a tool layer, so agents can actively query and interact with the ontology at runtime.
To build and maintain that layer, the authors introduce a builder agent for autonomous ontology construction, plus a self-evolution loop that continuously refines the ontology through attribution-guided typed edits. Those edits are accepted only after a backbone-conditional paired evaluation.
Experiments were run on three widely adopted data-agent benchmarks with four LLM backbones. The authors report that EvoOntology consistently outperforms strong baselines and existing semantic-layer approaches, which they say bridges the agent-data gap and enables more effective interaction with heterogeneous data.
The project is described as a universal agent plugin delivered as MCP, with integration cited for Claude Code, Codex, and other AI agents. Code is available at github.com/ruc-datalab/EvoOntology.