Companies that let engineers loose with coding agents are discovering they have no idea what the agents cost or which models they are calling. On 24 September Databricks announced a command-line tool, part of its Unity Gateway, that gives administrators one place to set which agents and models developers get by default, which tools and Model Context Protocol servers they can reach, and what they are allowed to spend.
The figures are all Databricks’ or its customers’ own. John Xing, chief technology officer at Concurrence, is quoted saying its coding agents have generated more than 61 billion input tokens across roughly 360,000 requests since rollout. Databricks says its own engineering team used the gateway’s tracing to find seven faulty tool integrations, and puts the saving at an estimated $1.2 million, a figure that combines wasted spend with lost productivity, so it is not a billing number. A cost-routing feature that sends easier tasks to cheaper models is credited with 35% savings on an internal coding benchmark.
Plumbing rather than news, but the direction is real: agent spending is becoming something finance departments ask about, and vendors are moving to sell the answer.