How Databricks Manages Its Own Coding Agent Spend

Rohit Agrawal and colleagues on a governance problem rather than a cost-cutting one. Thousands of Databricks engineers use coding agents every day, mixing Claude Code, Codex and Cursor, often several at once, and the spend is now one of the fastest growing line items in R&D. Their starting position is deliberately permissive: allow engineers to spend unimpeded by approvals and escalations, because AI leverage is the point.

A single monthly limit failed on both ends. It was slow to catch a runaway loop, since an automation could burn through a month of budget in an afternoon, and it created an approval queue: at their scale, somewhere between 500 and 1,000 engineers were hitting the limit every month. So they split it in two. A small daily limit, auto-resetting, catches runaway spend and can be raised by the engineer alone by acknowledging in Slack that the spend is intentional. A larger monthly limit handles real growth and needs manager approval, granted in tiers (2x, 5x, unlimited), scoped to a project and reverting automatically. The two are coupled by a fixed ratio and whichever binds first wins.

The dollar figures in the post are illustrative rather than their real numbers, a $500 default monthly limit with manual review past $2,500. In their actual deployment an engineer spending evenly across the month never trips the daily limit at all, because the monthly budget spread over working days sits below the daily threshold. The daily limit is there for anomalies, not for pacing.

Two details worth keeping. The self-serve Slack acknowledgement is a deliberate liveness check, since an unattended cron job cannot click a Slack button, so the human-in-the-loop step distinguishes intentional work from a loop nobody is watching. And budgets do not roll over: last month’s big push does not carry over as this month’s headroom. Routing everything through one gateway is what makes any of this enforceable across tools. Pairs with their earlier post on benchmarking coding agents on their own codebase, which covers which agents to buy; this one covers how much to let people spend on them.