Great writeup - I'm also facing the same problem with Genie Code.

1. This billing attribution gap isn't specific to Genie Code, it shows up anywhere an agent/orchestration layer spins up ad-hoc compute (serverless one) on your behalf. Eventhough you can control this via budget policy. 

2. It is quite frustrating to not able to control the underlying model of Genie Code, I'm not even sure if Databricks is doing semantic model routing to cheaper model when the task is easy or not.

3. I'm also sensing that for complex task, Genie Code tend to perform this weird loop that caused so much money.

4. I wish they have a usage view such as /usage, /context of Claude code, PI, or KIRO.

The only 2 ways I'm able to control the cost of Genie Code is

1. Via serverless budget policy.

2. and the account level budget for Genie.