AI Agent Cost Observability Becomes the Next Data Platform Fight

Agents moved from feature to workload

Snowflake and Databricks held their 2026 flagship conferences in San Francisco within about three weeks of each other. The clearest takeaway is that both companies now treat AI agents as core infrastructure rather than add-ons. One analysis of the two events noted that the platforms began from different places, the data warehouse and the lakehouse, yet landed on nearly identical product bets 5.

The scale numbers explain the urgency. Databricks expanded Agent Bricks, introduced Genie One, and said more than 100,000 agents have been built on its platform, processing over a quadrillion tokens a year 5. Snowflake rebranded Snowflake Intelligence as CoWork and Cortex Code as CoCo 5. It also introduced Cortex AI Gateway 2, which it framed around two barriers to enterprise adoption: securing agents and giving organizations centralized visibility into AI consumption costs 4.

This creates a spending problem that neither vendor's earlier billing models were built to show clearly.

The spend gap both vendors acknowledged

Snowflake described the issue directly. Agents that work across enterprise data, apps, and platforms drive growing AI consumption across models and workloads, and organizations often have no central way to govern that usage 4. The company argued that traditional security architectures were not designed for this kind of cross-system agent activity, and it called for a new security and cost model 4.

PointFive, a cloud cost firm, put the financial stakes more bluntly. It expects warehouse and lakehouse bills to split in two. Agent-driven compute was not a meaningful line item a year ago, and the firm projects it will be a top-three spend category for many teams by the end of the year, based on the adoption curves both platforms cited 5. Its main warning is about timing. Once agent calls blend into an analytics baseline, separating them after the fact is genuinely hard 5.

That forecast comes from a company that sells cost tooling, so it should be read with that in mind. The underlying point still holds up. When a new workload category grows quickly inside an existing bill, the window for clean attribution is short.

How the platforms are responding

Databricks presents Agent Bricks as a single control point for governance and monitoring. The company says it can track every agent, MCP server, model, and tool with clear ownership and end-to-end permissions 1. It says it captures every interaction, tool call, and model invocation without code changes, and evaluates quality continuously using LLM-as-judge, custom metrics, and human feedback 1.

On the evaluation side, Databricks says Agent Bricks, built on MLflow 3, automatically generates task-specific evaluation datasets and custom judges. It then searches across optimization techniques to raise quality, including prompt engineering, fine-tuning, reward models, and test-adaptive optimization 3. Synthetic data generation is meant to remove the need for manual labeling 13.

The framing differs between the two. Databricks leads with quality and governance, emphasizing measurement, evaluation, and permissions 13. Snowflake's gateway puts cost oversight next to security as a headline feature 4. Both are responding to the same sprawl, but they stress different halves of it.

Third parties fill in the layers

The surrounding ecosystem is moving quickly too. On June 15, 2026, Monte Carlo announced support for Agent Bricks. The company extended its data observability into the agent layer, so customers get a continuous view across the Databricks stack 2. Monte Carlo already monitors Delta Lake tables for freshness, schema drift, and volume anomalies, and tracks lineage across Lakeflow 2. Adding agents connects data health to agent behavior. That link matters because an agent fed stale or malformed tables can fail quietly while still running up compute.

Taken together, the market appears to be splitting into layers:

• Platform-native governance and evaluation: Agent Bricks, Cortex AI Gateway
• Data-quality observability that now reaches agents: Monte Carlo
• Cost attribution specialists: firms like PointFive, warning that teams need to separate agent spend early

What to make of it

The tooling is arriving quickly because the vendors themselves have identified the problem. When the platforms selling agent compute also ship features for monitoring and capping that consumption, it signals that customers are already asking hard questions about their bills.

The practical lesson for data and FinOps teams is to tag and attribute agent workloads now, before they mix into baseline usage. Built-in evaluation and observability help with quality and governance. Whether native dashboards will satisfy finance teams or open a lasting market for independent cost tooling remains an open question. For now, there is clearly a gap, and many companies are racing to fill it.