Internal Data Agent

This is an example recipe, not a core Little Harness concept. Use it when an agent needs to query data, write large results to files, inspect them with scripts, and return a compact answer.

Shape

const harness = createHarness({
  host: localHost(),
  model,
  system: "Analyze internal data. Prefer file-backed analysis for large results.",
  tools: { queryWarehouse, getMetricDefinitions },
  skills: [skill("./skills/data-analysis")],
});

Artifact-First Queries

Warehouse tools should be allowed to write large results to artifacts.

{
  preview: "18,204 events returned for the selected cohort.",
  artifactPath: "/artifacts/warehouse/turn_005/events.parquet",
  bytes: 8910231
}

The agent can then run focused analysis:

python3 /session/scripts/cohort_summary.py /artifacts/warehouse/turn_005/events.parquet

Keep Permissions In Tools

The database tool remains the permission boundary. The agent should not receive direct network or credential access just because it has a shell.

Good tool descriptions say:

• what tables or metrics are available
• whether the tool can mutate data
• what row limits apply
• where large outputs will appear

Little Harness manages the session workspace and artifact path. It does not protect the database from an expensive query. Put timeouts, row limits, read-only credentials, tenant filters, RLS, allowlists, and validation in the database tool or the backing data service.

Persist Only Derived Notes

For analytics agents, Persistent Dirs should usually store metric definitions, prior analysis notes, or team preferences. Raw query outputs belong in artifacts or source systems.