Autonomous BigQuery cost management, in three layers
Recommendations age fast; a Monday rollup doesn't wait for your review meeting. Tend runs the loop itself: forecast, policy, change, verification, in minutes.
Last updated
Tend runs on Google Cloud
Tend runs its agent reasoning layer on Gemini, which handles planning and tool use. See Google'sGemini function-calling documentation.
Google Cloud provides Identity and Access Management (IAM), audit logging, and network controls; each customer must review the configuration for its environment. Tend usesModel Armorto screen model inputs and outputs. Slot Agent records a trace for every applied action.
Proprietary demand forecasting
The models generate probabilistic demand curves at three horizons: minutes, hours, and days. Slot Agent uses these curves to set BigQuery reservation ceilings, BI Engine Agent uses them to warm tables, and Cost Agent uses them to revise spend forecasts.
Because the models train on your environment and not on generic cloud benchmarks, they detect patterns that generic tools miss. Examples include Monday morning rollups from the finance team, the extract, transform, load (ETL) job that runs only at quarter close, and the dashboard that spikes when a board package goes out.
Policy engine with confidence gating
Forecasting shows what demand is coming. The policy engine solves for the cheapest action that satisfies every business constraint at once.
The engine converts each forecast into a concrete action, such as raising or lowering a reservation ceiling, and solves it against the full constraint set: hard floors, ceilings, daily budgets, schedule profiles, and performance targets.
Tend logs every action with the forecast that produced it, the constraint set that was active, the confidence score for the recommendation, and the guardrail checks that the action passed before it applied. If the system or you rolls back a change, that signal feeds directly into how the engine weights future decisions.
When confidence drops below the threshold, the agent holds the action and alerts you instead of acting. You remain in control when uncertainty is high.
Multi-agent execution on Google Cloud for BigQuery
Slot Agent executes approved BigQuery reservation changes today; it does not only suggest them.
BI Engine, Query, and Cost agents each have a defined scope, toolset, and decision authority. A supervisor arbitrates between agents when their actions touch the same reservation.
The orchestration layer runs on Google Cloud with Gemini at the reasoning layer. Model Armor is configured to protect model inputs and outputs. Slot Agent records an auditable trace for each applied action.
Every live Slot Agent decision passes the same gate before it reaches a reservation. The gate includes freshness, drift, circuit-breaker, and convergence checks. If a check detects a problem, Tend restores the baseline and continues verification.
Autonomous tuning from day one
Traditional cost tools operate on a recommendation cycle that you measure in hours or days. By the time a recommendation surfaces, the workload has already changed. Tend closes that gap.
Tend runs forecasting, policy selection, and agent execution continuously on Google Cloud, not on a fixed schedule. As a result, Tend evaluates and acts in intervals that you measure in minutes, across your connected reservations. Savings appear as the models build history on your specific environment. For more information, see How we measure savings.
See how this applies to your warehouse
If you want to preview savings, connect a warehouse in read-only mode. Tend runs the forecasting and policy models against your reservation history for 30 days and shows you where savings are before you enable an automated action.