A shared agent ontology needs a change board, not just a graph
Neo4j’s Context Engine packages ontology, runtime and memory for agents. The harder control question is who may change a shared definition and how consumers are retested.
What happened
Neo4j announced Context Engine in private preview as a shared knowledge layer with business ontology, context runtime, studio and graph-native memory components.
Why it matters
When several agents depend on one business definition, a small semantic change can alter retrieval, policy and action across workflows.
Neo4j's 20 October announcement describes Context Engine as a private-preview package with Business Ontology, Context Build and Runtime, Context Studio, and Memory and Learning. The vendor says a shared layer can help agents reason over enterprise knowledge. Private preview and product description do not establish production reliability, accuracy or governance outcomes.
TechTarget independently confirms the components and places the release in a broader context-platform trend. It quotes analyst William McKnight saying the current layer is read-only and lacks native change-data-capture connectors and bidirectional write-back; he also calls for independent throughput, latency and total-cost benchmarks. The useful procurement question is therefore not whether graphs make agents trustworthy. It is whether the organisation can govern changes to meaning across every consumer while testing unproven operating boundaries.
Create a semantic change record for each concept, relationship, source mapping and policy binding. Name the owner, reason, affected data, agents and workflows, backward-compatibility decision, test pack and rollback. Version the ontology and runtime together so a result can be traced to the meaning in force at execution time.
Before promotion, replay representative retrieval tasks and the downstream actions performed by consuming workflows, including ambiguous entities, stale relationships, deleted records and conflicting sources. Those action tests evaluate the connected workflow; they are not a claim that Context Engine itself writes back to source systems. Compare both answers and resulting actions. A change that improves one agent can break another consumer or widen access by connecting previously separate data.
The counterargument is that governance slows iteration. Use risk tiers: descriptive concepts can follow a lightweight review; concepts that determine entitlement, payment, employment or customer treatment require domain and policy approval. Preview speed is not a reason to let shared semantics become an unowned production interface.
This extends the newsroom's skills-lineage contract. The immediate action is to require an owner, impact list, regression pack and rollback for every ontology change before connecting a second production agent.