Open Data Infrastructure
Agentic AI Query Budgets in Open Lakehouse Systems
How agent query budgets connect cost control, workload isolation, policy denial, retries, and evaluation traces in open lakehouse systems.
Agentic AI changes lakehouse cost control because the person asking the question may not be the thing generating the queries.
Agents spend through queries
An agent can plan, retry, broaden a query, call a tool again, and run evaluation probes. In an open lakehouse, that behavior can touch shared serving clusters, open table scans, catalogs, and downstream workloads.
StarRocks query queues and resource groups are examples of serving controls that can limit concurrency and allocate resources. Policy engines such as OPA can evaluate access decisions. Evals can show whether budget limits damage answer quality.
A budget is a control surface
A query budget should define how many requests, how much scan volume, which workload class, which retry behavior, and which escalation path an agent can use for a task. It should also define what denial looks like.
The budget must be visible to the agent platform and the data platform. Otherwise, the model may keep trying while the serving tier quietly absorbs the cost.
Core idea: Agent query budgets are where cost control, policy, and reliability meet.
The ODI operating model
Open Data Infrastructure should attach query budgets to catalog identity, data product contracts, workload queues, and eval traces. That gives teams a way to tune the budget without guessing whether lower cost broke the task.
For adjacent context, read agentic AI write paths and human review, StarRocks query queues, and retrieval governance.
What breaks first
- The agent runs under a generic service identity with no per-task budget.
- Retries multiply cost after policy denials or queue waits.
- Serving operators see cluster pressure but not the agent task that caused it.
- Eval traces measure answer quality but not query cost or denial behavior.
Questions to ask
Ask how budgets are assigned, enforced, and reviewed. Ask whether query traces include task, identity, data product, workload class, cost estimate, denial reason, and eval outcome.
If an agent can ask the lakehouse to do work, the lakehouse needs the right to say enough.
Sources to start with
These primary sources anchor the technical claims in this guide.
- StarRocks query queues documentation
- StarRocks resource groups documentation
- Open Policy Agent documentation
- OpenAI evals documentation
- NIST AI Risk Management Framework
An agent without a query budget is a cost center with a prompt.