Something strange is happening inside enterprise data infrastructure. While most companies are still running pilot programs, AI agents have quietly taken over the work of creating databases. On Neon, the serverless Postgres database that powers Databricks Lakehouse, AI agents now generate 80% of all new databases. Two years ago, that number was 0.1%.
That jump — from near-zero to dominant infrastructure player in 24 months — is the real story in Databricks' new 2026 State of AI Agents report, released this week. The data comes from aggregated telemetry across more than 20,000 Databricks customers, including 60% of the Fortune 500, making it one of the broadest real-world snapshots of enterprise agent adoption available.
The headline finding: only 19% of audited organizations have deployed AI agents at scale. But that aggregate number obscures a more complicated picture. In specific workflows, agents aren't experiment — they're already foundational infrastructure.
The Database Shift
The most striking shift is in database provisioning. Agents now create 97% of all database branches on the Neon platform. Databricks calls this a consequence of "vibe coding" — natural-language-driven application development where non-technical users describe what they need and agents handle the infrastructure. The result is ephemeral database environments spun up in seconds rather than the days or weeks traditional provisioning requires.
This is not a future-state scenario. The 80% figure comes from internal telemetry, not surveys. The infrastructure exists. It's running.
Multi-agent workflows tell a similar story. Usage on the Databricks platform grew 327% between June and October 2025 alone, driven by new orchestration features. The most common pattern isn't a single agent — it's a Supervisor Agent coordinating subagents across specialized domains. That Supervisor Agent accounts for 37% of all Agent Bricks usage.
The Governance Gap
The same data that shows agents succeeding also exposes a gap. Most organizations are deploying governance after adoption, not before. Craig Wiley, senior director of product for AI/ML at Databricks, put it directly: "Once organizations have built enough that it starts to get nerve-wracking, they start to worry about governance."
The correlation is stark. Companies using Databricks' AI governance tools deploy 12 times more AI projects to production than the average firm. AI Gateway, the company's governance product, grew sevenfold in nine months. And organizations using evaluation tools — structured tests that verify agent outputs — achieve nearly six times more production deployments.
Evaluations remain "existential," Wiley said, when agent output affects financial or reputational risk. They're also underused, partly because building good evaluations is hard and the industry hasn't prioritized them.
What This Means for Infrastructure Teams
The speed of agent-driven provisioning is outpacing how IT departments think about their role. Database administration built around human provisioning cycles doesn't map well to infrastructure that spins itself up on demand. Wiley expects organizations to need "a new way of thinking about how they govern the organizations and the business units they work with."
The agent layer isn't just automating tasks. It's changing the shape of the infrastructure stack itself — and the teams that manage it.
Primary sources: Databricks 2026 State of AI Agents report | SiliconANGLE coverage
