ExtraltExtralt

Using the dashboard agent

The dashboard agent helps operate supported workflows and answer bounded data questions. It is not an unrestricted autonomous operator.

What it can read

The agent can list or inspect organization-scoped:

  • Robots;
  • Extract Runs;
  • Enrichments;
  • Schedules;
  • Imports.

It can also generate and execute a small persisted query over Captures, Items, or Changes. The generated result contains at most 100 rows and opens in the matching dashboard AI view.

What it can propose

The agent can prepare approval plans for:

  • one new Extract Run using an existing Robot;
  • one new Extract Schedule using an existing Robot;
  • one Enrichment for an existing Extract Run.

The plan includes the exact selected resource, budget, URLs, cadence where applicable, and whether Enrich follows Extract. The dashboard shows the plan before execution.

Approval boundary

Preparing a plan does not run it. State-changing steps execute only after the user approves the plan in the dashboard. Rejecting or superseding the plan does not consume extraction or enrichment credits.

The agent does not:

  • create, build, or ensure a Robot;
  • select a Robot silently without confirming its name;
  • create or retry Imports;
  • query Products, Variants, Listings, Offers, Reviews, or Stores through its generic query tool;
  • execute arbitrary ClickHouse SQL;
  • bypass organization isolation or dashboard approval.

Ask precise questions

Use Capture for raw source records, Item for enriched page records, and Change for observed changes. “Product” in a casual question can refer to either a Capture or Item; say “canonical Product,” “Variant,” “Listing,” or “Offer” when you need the final Extend entity.

Examples:

  • “Which Captures from this Run have no brand?”
  • “Group these Items by category and brand.”
  • “Show price Changes from the last day.”
  • “Prepare a 100-Capture Run with the confirmed Nike US Robot, then Enrich up to 100 Captures.”

For final current-market or Variant questions, use Explore. The current agent does not yet expose a dedicated Explore analysis tool.