Alibaba Cloud's cloud-native community published on October 26 an introduction to UnifiedModel, an open-source semantic layer that describes enterprise assets, runtime data and system relationships as a queryable object graph. It rests on three primitives: sets for entity, metric, log, trace, event and runbook classes; links between them; and fields that carry type, unit and query mapping. The article splits a stable definition layer from runtime entities that are written and expired as the system runs. Its argument is that agents fail on production questions not because a single tool reports a wrong number, but because nothing tells the model which object each number belongs to, how objects depend on each other, or what changed upstream. In the authors' DataAgentBench experiments, four flagship models each gained 10 to 20 percentage points once the layer was in place, with GLM-5.2 reaching 50.2% pass@1.
Our takeThis is the schema argument in agent clothing: retrieval supplies fragments, an object graph supplies the join path. The design choice worth noting is that fields carry units and semantics, which is what lets a generated query be checked rather than merely written. The benchmark is the authors' own and the layer was built by the same team, so the 10 to 20 point range needs outside replication. The concrete test is whether an agent can localise a root cause without being handed the topology in its prompt.