Genie One analyzed the trace data and produced a ranked list of recurring tool errors, recovery turns, and their token and wait-time costs, forming the basis for targeted fixes.

The core insight isn’t just better error messages, but designing MCP tools to be forgiving of LLM-style inputs and under-specified signatures to prevent token waste.

Team used Unity Gateway tracing to capture OpenTelemetry traces of all MCP tool invocations, enabling attribution of costs to specific tools, errors, and agent sessions without extra instrumentation.

Databricks reduced wasted AI agent spend by $1.2 million per year after identifying and fixing seven small tool-server bugs within one hour.

A continuous monitoring loop is established: observe via Unity Gateway, query with Genie One, identify and fix, then repeat to maintain cost efficiency as agent usage grows.

Fixes were straightforward to implement across tool servers once buggy patterns were identified, with the entire find–quantify–fix loop taking about an hour.

Example bugs include Jira key errors and Google Drive/Docs field errors, with quantified impact such as 535 Jira errors per day costing $87K in tokens and 4,850 hours of wait time annually.

Databricks promotes Unity Gateway and Genie One as a repeatable approach to monitoring, diagnosing, and reducing AI agent costs, encouraging teams to start with the unified trace table (Beta).

Root causes often lie in under-specified MCP tool signatures, which tempt models to guess parameters; the fix is to design tools that tolerate and absorb such guesses rather than crash.