Databricks Unity Gateway Cut $1.2M Agent Waste

Databricks traced roughly $1.2 million a year in wasted spend to seven MCP bugs. Engineers fixed them in about an hour using Databricks Unity Gateway and Genie One.

Companies in this story

StartupHub profiles of the companies this article names, with funding and a one-liner from our database.

Google

$9.1B

Global technology leader in search, advertising, cloud, AI, and consumer electronics.

Bloomberg L.P.

$42.6B

Global financial, software, data, and media company providing real-time information and analytics.

HumanX

$23.0B

A company that organizes premier AI conferences and publishes data-driven reports on the AI economy.

Founders Fund

$19.9B

Venture capital firm investing in revolutionary technologies and ambitious founders tackling big problems.

The trouble sat in internal Jira and Google Drive and Docs tool servers. No remote exploit was needed. Any agent that sent a plausible JSON type could kick off the loop.

How the attack actually works

Unity Gateway spits out an OpenTelemetry trace for every MCP call, capturing tool name, arguments, errors, tokens and latency. Genie One queried that table in plain English and ranked failures by volume and cost.

The worst offender logged 535 failures a day. It triggered when the model sent an array for the fields parameter instead of a comma string. The server called .split() on a list, threw "'list' object has no attribute 'split'", and the agent retried for 12 turns on average. Picture a vending machine that only takes exact change and jams. The user does not walk away. They shake it, try again, burn time and money.

Why this matters, and what it does not fix

The pattern is quiet. Tasks still complete, so aggregate dashboards read it as healthy usage growth. At Databricks, it added up to 1,409 errors a day, $499K in annual token waste and 12,023 hours of wait time, with one Drive error hitting 49.6% of drive_file_get calls.

Recovery cost tracked error quality. Clear messages like "find_text and replace_text required" recovered in 4.6 turns with a 14% repeat rate. Cryptic tracebacks took 12.1 turns, and misleading kwargs hit a 50% repeat rate. Builders should coerce inputs and return actionable errors, and keep tracing on. Without the gateway on the call path, there is nothing to query.