
Why it mattersA team paying a coding-agent bill now has a reader that ranks the three biggest sources of waste.
A coding agent that retries a broken tool call ten times still prints a successful result at the end. The user pays for every retry. Databricks has released a product that reads an agent's trace log and names the retries that cost the most. Adam Gurary, Alkis Polyzotis and Vishnu Kannan announced Unity Gateway Insights on the Databricks blog. The tool is a report that reads agent traces in Unity Gateway. It is open to teams already using the Gateway for coding-agent traffic.
Databricks writes that Insights ranks three categories of waste by cost. The first is recurring MCP errors. The second is simple tasks running on expensive frontier models. The third is repeated workflows worth turning into shared skills. The tool runs across the last day, seven days, or 30 days. It lets a user add up to ten custom questions per report. Every finding links to an example session.
How Databricks says it works
Genie writes the SQL and reads the sessions. Genie Ontology gives the answers your own project names and business terms. A trace from Claude Code, Codex or GitHub Copilot already goes through Unity Gateway. The team has to route that agent traffic through it first. Insights then sits on top of that data. The user opens it at AI Gateway, then Observability, then Insights. There is no new install step.
The $1.2M number Databricks puts behind it
Databricks links to a second post. In it, the company says the trace analysis recovered an estimated $1.2M in costs a year inside Databricks itself. The breakdown Databricks gives is $499,000 in wasted tokens, plus 12,000 hours of engineers waiting on agents. Both figures are Databricks' own, and they come from seven small bugs in its own MCP servers. Databricks does not state whether another customer has reproduced the finding.
A reader applying this on their own fleet should treat the $1.2M as Databricks' own saving. Nobody else has tested the figure yet. The useful part of the post is the method. Three waste categories, ranked by dollars, each with a sample trace. That is a reasonable starting shape for a team with heavy agent traffic.
Where this fits in the market
Many AI observability products sell on trace replay, cost attribution and eval scoring. Insights sits close to that space. A buyer should ask one question. Can an existing tool read the same Unity Gateway logs? The lock-in question matters more here than the feature list. The cost signal lives in the logs themselves.
Databricks says the product is available now. The post does not state a price. It does not state a preview label either. A team already paying for Unity Gateway should check the Observability tab on their account.
Source
Introducing Unity Gateway Insights by Adam Gurary, Alkis Polyzotis and Vishnu Kannan on the Databricks blog.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.