When to use this profile
Use a local database snapshot when the agent needs relational analysis without access to the live production database. Open the supplied SQLite file in read-only mode, inspect the schema and run bounded queries. Return aggregates and a compact sample so the model can reason about the answer without receiving an entire customer table.
Prepare a repeatable environment
- Create a data export with only the rows and columns needed for the task.
- Open the input through SQLite's read-only URI mode and write derived tables to a separate file.
- Set query deadlines with a progress handler and impose a maximum result-row count.
python /work/query_database.pyReturn results the agent can use
- The SQL query and structured results with column names.
- CSV exports for approved aggregate tables.
- A report of rows scanned or returned where the application can measure them.
Query a fixture database, verify a known aggregate and attempt a write that should fail. Include an expensive query to confirm the deadline interrupts it.
Resources and boundaries
Start with 2 vCPU and 4 GB of RAM, then measure peak memory and task duration on a representative fixture. These are workload planning values, not a benchmark or a provisioned configuration. Use the sandbox cost calculator to estimate running time and retained snapshots.
- Read-only SQL can still be expensive through joins, sorts and recursive queries.
- Disable loading extensions and avoid attaching arbitrary filesystem paths.
- A sandbox boundary does not replace data minimization for sensitive database exports.
Keep model inference separate from this execution profile. A hosted model or gpuOS can decide the next action while the CPU environment runs it. The quickstart describes the account workflow and the proposed runtime contract.