Databricks data shows multi-agent AI adoption accelerating

By Olivia Hart

Databricks telemetry shows multi-agent workflows on its platform grew 327% in four months during the second half of 2025. Databricks says the dataset covers more than 20,000 organizations and documents a production shift from single chatbots toward coordinated multi-agent systems.

How quickly are enterprises adopting agents?

Databricks telemetry includes more than 60% of the Fortune 500, according to its analysis. The company's marketing materials cite 70%, a discrepancy the source leaves unexplained.

A G2 survey of more than 1,000 B2B decision-makers, reported by late 2026, found that 66% of organizations had agents in limited or scaled production. Another 86% had purchased agent software or services during the prior year, up from 69%, although the source does not specify the comparison period.

Databricks reports the following platform usage:

• Agents on Neon, Databricks' serverless Postgres platform, create 80% of databases and 97% of branches used for testing and development.
• Customer-experience work, including support, onboarding, personalized outreach and market intelligence, accounts for 40% of the top 15 use cases.
• The Supervisor Agent, which coordinates specialized agents, represents 37% of usage on Databricks Agent Bricks.

Takeaway: Enterprises are deploying agents at scale, but breadth of adoption does not yet demonstrate mature, autonomous operations.

What is adoption costing in oversight?

G2 found that 45% of buyers said oversight and rework consumed more than a quarter of the productivity gains they expected. G2 calls that cost the “Distrust Tax,” a label that captures how supervision can absorb the value automation promises.

Thirty-one percent of G2 respondents said employees spend six or more hours each week overseeing agent work. Only 7% expressed comfort with full autonomy, leaving most organizations dependent on people who review, redirect or repair agent decisions.

Anthropic's survey of more than 500 U.S. technical leaders, conducted in late 2025, found that 57% of organizations used agents for multi-step workflows. Only 16% reached cross-functional or end-to-end processes, while coding agents were described as nearly ubiquitous.

Takeaway: Agent deployments are spreading faster than organizations can measure their productivity, govern their actions or reduce human supervision.

When does multi-agent architecture make sense?

In an October 22, 2026 article, Anthropic's Applied AI team said that, across several enterprise deployments, “a single primary agent equipped with skills and on-demand context frequently matched or exceeded multi-agent pipelines on quality while reducing cost and latency.” In one customer comparison, the single-loop configuration ran roughly twice as fast.

Anthropic published a migration kit to help teams evaluate consolidation. It also said multi-agent designs remain appropriate when tasks are genuinely parallel and independent, when isolated contexts are required, or when tool libraries become large enough to degrade selection quality.

The evidence for consolidation remains limited. Anthropic did not specify how many deployments the comparison covered or which industries participated, so the result does not establish that one architecture should replace the other.

Takeaway: Multi-agent systems should solve a defined coordination problem, rather than become the default response to rising demand for agents.

Where do security and governance fall short?

The December 2025 OWASP Top 10 for Agentic Applications lists goal hijacking through indirect prompt injection, tool misuse, identity and privilege abuse, and cascading failures among its principal risks. The list treats agent systems as a distinct security category rather than a conventional application stack.

A BlueRadius dataset of 21 publicly documented AI-agent security cases found excessive privilege in 76% and prompt injection in 71%. Every case either bypassed human approval or lacked human approval in the workflow, although only 19% were confirmed real-world incidents.

The remaining cases were disclosed vulnerabilities or red-team demonstrations rather than confirmed breaches. That distinction limits the dataset's use as a measure of production prevalence, even though its failure patterns remain relevant to system design.

Gartner forecasts that an average global Fortune 500 enterprise will use more than 150,000 agents by 2028, up from fewer than 15 in 2025. Only 13% of organizations currently believe they have adequate governance for that scale, according to Gartner.

Takeaway: Forecast agent volumes are rising far faster than confidence in governance, increasing the cost of weak permissions and unmonitored actions.

Do evaluation tools predict production success?

Databricks reports that organizations using evaluation tools put nearly six times more AI projects into production. It also says organizations using AI governance put more than twelve times more projects into production than the average.

Those findings are associations in platform telemetry, not controlled evidence of causation. Organizations that adopt evaluation and governance tools may also have larger teams, bigger budgets, more mature data or broader project portfolios.

Takeaway: Better tooling accompanies higher deployment rates, but the telemetry cannot prove that the tooling alone causes production success.

What remains unproven?\n

The evidence base contains several boundaries that its headline percentages do not cross:

• Databricks telemetry reflects its customer base and is not statistically projectable to all enterprises.
• The G2 and Anthropic surveys cover different samples and time windows, preventing their results from forming one market-wide estimate.
• BlueRadius selected publicly documented cases and cannot establish how common agent-security failures are across enterprise deployments.
• Anthropic's single-agent comparison provides no sample size or industry breakdown.
• No source provides a longitudinal view of whether multi-agent architectures sustain value once governance and cost constraints are applied.

Takeaway: The next test is whether enterprises can convert rapid deployment into sustained output after accounting for rework, supervision, latency and governance.