Slot Agent
Tend's Slot Agent runs on Google Cloud with Gemini for reasoning. Instead of showing recommendations for you to approve, the agent acts only when you enable auto-apply. Every five minutes, the agent reads reservation telemetry, runs its forecasting models, and sets a new ceiling through the approved reservation-write path. If confidence drops below the threshold, the agent holds the ceiling and alerts you instead of acting. See Google'sBigQuery reservations documentation.
Forecast demand from your workload history
The ceiling is only as strong as the demand model behind it. The Slot Agent runs two models against your workload history instead of a generic benchmark. The five-minute nowcast model predicts near-term slot demand and enables a fast reaction. The planning model forecasts 24 to 48 hours ahead, so the ceiling accounts for known peaks such as quarter close, Monday rollups, or recurring extract, transform, load (ETL) windows.
If planning data is missing or stale, Tend does not estimate. The agent keeps the hard ceiling that you configure instead of acting on incomplete data.
How the policy computes the ceiling
Most ceilings use one static number, sized for the worst day and unused on other days. Tend computes a live ceiling continuously and rounds it up to the nearest increment that your reservation allows. You can tune the forecast band. A wider band protects against bursts, and a narrower band saves more money.
Business constraints
Every warehouse has different risk tolerance by time of day. Tend lets you set that tolerance explicitly instead of using a single default.
Use schedule profiles to apply different settings for business hours and off-hours. For example, use tighter limits during the day and looser limits overnight. Use a floor to set the lowest value that the ceiling can reach. Use a daily budget to limit how far Tend can move the ceiling in a day.
Use a performance target to balance cost and speed. For example, tell Tend to keep average job runtime under one minute or to keep queuing time under 30 seconds. Tend then finds savings within that boundary.
If you need to fix capacity for a launch or an incident, use an operator override. The override includes a time limit and expires automatically, so it does not remain active after you no longer need it.
Guardrails
Every change passes the same gate before it reaches your reservation. Tend checks freshness, checks for drift, opens a circuit breaker after repeated failures, and verifies convergence to confirm that the change takes effect. If a check fails, Tend reverts to your baseline automatically and continues verification for several minutes before it marks the incident as resolved. Every applied action is recorded in the activity log with the before and after delta. For more information about measurement, see How we measure savings.
Static ceiling compared with Tend
| Approach | Ceiling | Change control |
|---|---|---|
| Static | One ceiling sized for the worst expected peak | Operator changes it manually |
| Tend | Recomputed every five minutes from forecast demand | Acts only inside approved policy and restores the baseline if a check fails |
Tend aligns the configured ceiling with forecast demand instead of leaving one worst-case ceiling in place. Actual reservation spend still depends on your BigQuery pricing and commitment model.