AI Pulse · Under the Radar

Databricks says it found a costly problem hiding in the connections between its AI agents and internal tools. Seven small MCP servers, used to access systems such as Jira, Drive and Docs, were accepting almost any plausible-looking JSON request. When those requests failed, the agents silently retried them.

That sounds minor. It isn’t. Databricks estimates the behaviour was wasting about $499,000 a year in tokens, while agents waiting on those retries added roughly 12,000 engineering hours a year in lost time. The company says it identified and fixed the issue in an hour using its own tracing tool. These are vendor-published figures, not independently audited, but the underlying lesson is solid. Databricks’ report is a useful reminder that AI costs often sit in failed handoffs, not just in the model price.

If you’re running agents that call internal systems, don’t only track total spend. Track failed tool calls, retries, response times and malformed requests by connector. A tool that “mostly works” can quietly create a large bill and make staff think the AI itself is slow or unreliable. Set clear validation rules at the tool boundary, fail visibly when requests are wrong, and put limits on retries.

This is the kind of thing we build into an AI command centre, because owners need to see where agents are spending money and time, not just whether someone has turned them on.