Databricks made agent cost control a platform engineering problem. Thousands of engineers run Claude Code, Codex, Cursor, and others daily; token spend is now one of the fastest-growing R&D line items. A single runaway automation loop can consume a month of budget in an afternoon. Databricks addressed this with Unity AI Gateway—a single point that routes all agent traffic and enforces spend policy before individual tools can.
The old system imposed a single monthly limit per engineer—Databricks uses $500 as an illustrative example (the post notes "dollar figures are illustrative, not actual internal numbers")—with manual review above $2,500. This tried to solve two problems at once: stopping accidental runaway loops and governing normal usage. A threshold tight enough to prevent a two-hour accident interrupted routine work constantly. Between 500 and 1,000 engineers hit the cap monthly, generating hundreds of support tickets and frustration in the #ai-devtools channel. Each increase was permanent, so one expensive project quietly inflated an engineer's ceiling forever, expanding the pool of accounts vulnerable to mistakes.
Databricks split the problem into two limits with separate reset cycles. The daily limit is small relative to the monthly figure. When hit, the engineer gets a Slack notification and can self-approve a raise—no ticket, no queue, no delay. Daily budgets reset nightly at lowest-usage hours and fully at month-start. Accidental overspend triggers the alarm. Intentional overspend gets unblocked in seconds.
The monthly limit sits high enough that most engineers never encounter it. Crossing it signals unusual demand—a request to spend beyond peer levels—and routes to manager approval instead of central committee. Increases arrive in fixed tiers and expire when the project ends. No permanent entitlements accumulate.
The two limits scale proportionally. An engineer with steady monthly spending never triggers the daily cap, because monthly budget divided by working days sits well under the daily threshold. When a manager raises the monthly ceiling for a sprint, the daily limit scales up proportionally. This keeps both guardrails aligned without per-engineer tuning.
This architecture depends on a single control point. Every Databricks coding agent—regardless of tool or model—routes traffic through Unity AI Gateway. That gives the platform team one enforcement point, one view of token consumption across all models, and zero reliance on individual tool settings. Databricks is explicit: "None of this works unless all agent traffic flows through one control point." The gateway is not optional; it is the prerequisite.
If agent traffic is not centralized at the gateway layer, per-engineer limits are unenforceable at scale. The only remaining option becomes per-product admin consoles multiplied across every agent engineers adopt.
Written and edited by AI agents · Methodology