Your database's newest client is an agent, and the vendors just rebuilt for it
Cockroach, Neo4j, Vespa, and Starburst all shipped agent-facing features. The common thread: agents ask worse questions than people, and the data layer has to absorb that.
- Vespa says its full schema skill raised agent success on data-dependent tasks to 93%, from 9% without explanations, while cutting agent query volume by 75%. Self-reported m:…6067
- Starburst cites a VentureBeat survey finding 78% of respondents with agents on company data traced errors to missing or conflicting definitions. Self-reported m:…3506
- Neo4j's Virtual Graph is now generally available on Neo4j Aura on GCP, allowing zero-ETL knowledge graphs and GraphRAG systems. Self-reported m:…9141
- Cockroach Labs made bring-your-own-cloud available on Continuum, with clusters running in the customer's own cloud account at $1,000 per vCPU per year. Self-reported m:…3859
The agent as a client
Vespa's new agent-ready schemas start from a blunt observation: agents are bad at guessing what a database means. From version 8.786.55 onward, Vespa automatically generates schema skills for AI agents, exposed at an endpoint when configured, and developers can add explanation blocks to schemas to give context for fields and rank profiles. The explanations update automatically as the schema changes. The measured difference is large. In Vespa's agent-discovery evaluation, the skill without explanations achieved a 9% success rate on data-dependent tasks; the full skill achieved 93%. Using the full schema skill reduced agent query volume by 75% compared with no skill, and reduced agent model time by 30% and token usage by 40%. On the STaRK-Amazon dataset, the schema skill increased relevance by 12% and reduced token cost by 68%. m:…6067
Those numbers are Vespa's own, from its own evaluation, and they should be read as a vendor's best case. Even discounted, they point at the same conclusion every database vendor reached this week: an agent that cannot see the schema will burn tokens asking the wrong questions, and the fix belongs in the data layer, not in the prompt. m:…6067
Definitions are the bottleneck
Starburst makes the argument at the level of meaning rather than schema. Enterprise AI agents lack autonomy, Starburst says, because business definitions are siloed within individual platforms. Customer-relationship, billing, and data warehouse systems maintain distinct, conflicting definitions for the same business terms. The company cites a VentureBeat survey from October 2026 indicating that 78% of respondents with agents on company data traced errors to missing or conflicting definitions. Starburst proposes harvesting definitions from existing catalogs and semantic layers to create a unified context. The pitch is that an agent cannot be trusted with a number until the organisation agrees on what the number means, and that agreement has to live somewhere the agent can read. m:…3506
Graphs and databases follow
Neo4j's contribution is aimed at the retrieval side. Virtual Graph is now generally available on Neo4j Aura on GCP, allowing the construction of zero-ETL knowledge graphs and supporting the creation of GraphRAG systems. The company plans to release support for AWS and Azure soon. The idea is to let an agent query a graph built over data where it already lives, instead of waiting for a pipeline to copy it into a store the agent can reach. m:…9141
Cockroach Labs, meanwhile, shipped the deployment model that regulated customers keep asking for. Cloud bring-your-own-cloud is now available on Cockroach Continuum, with clusters running in the customer's own AWS, Azure, or Cloud account, and carrying the same service-level agreements as fully managed Mission Critical clusters. Pricing is $1,000 per vCPU per year, with the Cluster Virtualization add-on at $200 per vCPU per year, and Mission Critical includes a 30-minute P1 response. Cluster Virtualization is in public preview on AWS and Cloud, and clusters can be created from the Cloud Console, the Cloud API, or Terraform. The message is that agent workloads belong in the customer's account, under the customer's controls, not in a vendor's multi-tenant cloud. m:…3859
What to watch
Four vendors, four layers of the same stack: the schema, the definitions, the graph, and the deployment. The common bet is that agents become the heaviest database clients an organisation has, and that the vendors who teach their systems to explain themselves to an agent will win the workloads. The open question is whether these features actually reduce the token and compute cost of agent work in production, or whether they move the cost from the agent's prompts into the data layer's plumbing. Vespa's numbers say the savings are real in its own evaluation. The honest test will be whether other vendors publish the same kind of measurement, on the same kind of task, with the same definitions of success. m:…6067 m:…3506 m:…9141 m:…3859
What is still unresolved
- Whether agent-facing schema features deliver the same savings outside the vendors' own evaluations.
- Whether organisations converge on shared business definitions, or each platform keeps its own.
- How much of the agent data layer ends up in the customer's cloud account versus vendor-managed services.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.