PLANNED Nothing on this page is built yet.
Agents
Four agents, each built for one job: write a pipeline, check a table's quality, answer a question in SQL, or diagnose a failed run. None of them runs on its own, none of them spends without you seeing the cost first, and none of them shows you anything you could not already see yourself.
What each agent does
| Agent | Returns | Lands |
|---|---|---|
| Pipeline Agent | A pipeline change set: SQL models, the asset graph, tests, documentation, a format recommendation and its cost. | On the change set's page, under Changes. |
| Quality Agent | Expectations as a change set, anomalies by count, and a classification suggestion. | On the change set's page, and on the table's Quality tab. |
| Analyst Agent | A SQL statement with the tables, metrics and freshness it rests on. | In the SQL workspace, pending, never run. |
| Repair Agent | A diagnosis of a failed run, and a fix as a change set. | On the run's Diagnosis tab, and the fix under Changes. |
How a request works
You ask, an agent proposes, you decide. A request always starts with you asking for it; nothing sends, runs, submits, approves or applies on an agent's behalf. What an agent returns is never chat — it is one of four things you already know how to review:
- A pipeline or a quality proposal is a change set, reviewed and approved the same way as anyone's, on its own page under Changes.
- An answer is a SQL statement that lands in your workspace and runs only when you run it.
- A classification is a suggestion; you are the one who classifies the table.
- A repair is a diagnosis plus a fix, and the repair run itself is one you start.
What was checked and what was read always come before the agent's own account of its reasoning, and that account is always shown as the model's words, not yours.
What it costs, and when you see it
Before you send a request, you see the ceiling on what it may spend, your budget, and a measured typical cost for a request like it. Money is the only figure shown, in dollars, never a token count or a model-specific unit. A token count appears exactly once: on the receipt of a call that has already happened.
What it can see
An agent sees only what you could see yourself, under your own access, and never more. A source you cannot see is shown as a count, not a name. And an agent never puts a value on the screen that your own reading did not already allow: not a profile value, not a sample row, not a line of a log, not even the text of its own prompt. Where a value is withheld for this reason, the screen says so in one sentence, rather than showing nothing and letting you guess why.
Not built yet
No agent is built yet. Nothing here sends a request.