Artificial Intelligence

Databricks finds $1.2M in agent waste

Databricks launched Unity Gateway Insights after finding $1.2M in yearly agent waste in its own coding traces.

StartupHub.ai Staff

By StartupHub.ai Staff, reviewed by Daniel Singer, Editor

2 min read

Databricks Unity Gateway Insights token waste report view
Credit: Databricks

Databricks launched Unity Gateway Insights after finding $1.2M in yearly agent waste in its own traces.

Insights runs on live Unity Gateway traffic and uses Genie to rank token waste by cost, Databricks says.

Offline replays showed no performance gain moving from Opus 4.8 to Opus 5, so the team stayed on 4.8, Databricks says.

BenchLM comparison puts Opus 5 ahead of Opus 4.8 on coding at 66.9 to 55.8, so the saving is workload specific.

Opus releases also differ on thinking controls, with 4.8 rejecting non-default sampling and later versions tightening disable options, per comparison documentation.

Seven MCP server bugs caused $499K in wasted tokens and 12,000 engineer waiting hours, Databricks says.

Fixing those seven bugs took about an hour, Databricks says.

Insights compares each workstream share of sessions with its share of tokens, flagging a 10% session workstream at 35% of tokens first, Databricks says.

Each waste type is ranked by token cost with links to example sessions, Databricks says.

Teams can add up to 10 custom questions to a report, including whether a newest model is worth it, Databricks says.

Teams can ask Genie which MCP server wasted most tokens and who owns it to draft a bug report, Databricks says.

Insights rates task complexity in frontier model sessions and counts routine sessions for Smart Routing to cheaper models, Databricks says.

A workflow qualifies after three sessions or two engineers, and Insights drafts the shared skill prompt, Databricks says.

Teams launch Claude Code, Codex or GitHub Copilot with the ug CLI for automatic OpenTelemetry tracing, Databricks says.

Traces land in a governed Delta table in Unity Catalog in an open format any tool can read, Databricks says.

Genie Ontology maps traces to project names and business nuances for investigation, Databricks says.

Databricks SQL queries full trace history at millions of traces scale, Databricks says.

Only the table owner can query traces until admins grant access, so Insights reads in place without export, Databricks says.

It only sees teams already routed through Unity Gateway with the trace table set up.

Reports for the last day, 7 days or 30 days can be generated today in AI Gateway Observability Insights.

Mentioned in this story

Sources

Get the free daily brief every morning

Every morning the day's AI news, new AI startups and new funding rounds. Free and takes a few minutes to read, unsubscribe at any time.

Join 64,000 other subscribers.

TopicsDatabricksUnity Gateway InsightsUnity GatewayGenieOpus 4.8Opus 5MCPSmart Routing

StartupHub.ai StaffDaniel SingerWritten by StartupHub.ai Staff, Editorial teamReviewed by Daniel Singer, EditorAbout

StartupHub.ai Staff, Editorial team

The staff writers of StartupHub.ai, ranging from investment analysts to avid AI tool users, early adopters and critical enthusiasts. Backgrounds span engineering, business and the arts. We hold every piece to rigorous standards of research and review.

Daniel Singer, Editor

Daniel Singer is an AI expert who has been writing about, covering and building AI technology since 2018.

More from Daniel Singer

Test your Agent Readiness now

Paste your website. How well ChatGPT, Claude and the other agents can read and use a site, in about a minute.

Watch

The videos behind the news, one flip at a time

The feed mixes the videos from our articles with startup demos and you flip through them like TV channels for free with no signup.