Skip to main content
The Quickstart runs a task start to finish. This is what you build once that is no longer enough: a list that a person checks before anyone trusts it, and that gets corrected in place rather than regenerated. Each step links to the guide that covers its mechanics. What this page adds is the order, and the decision at each seam.

​
1. Declare the shape, don’t describe it

Start from a spec rather than a prompt. A reviewer can only check rows against a declared shape, and a prose description is not one. → Structured output for output.schema and schema_order → Writing task prompts for the per-row rules that set the bar Decision: if the columns aren’t known yet, you are still exploring. Use a prompt and this pipeline later.

​
2. Watch the first run, don’t poll it

For the first run of a new spec, stream it. You are not waiting for an answer, you are looking for the agent misreading the brief, which shows up in the tool calls long before it shows up in the rows. → Live messages for the SSE routes and event names Decision: once the spec is stable, switch to polling. Streaming is for when you are still learning what the task does.

​
3. Read rows with their sources

Fetch the rows and keep the cells rather than flattening immediately. Review is exactly the case where _sources earns its keep: a reviewer needs to see why a value is claimed. → Outputs and sources for cell shape and _sources

​
4. Correct in place

Send the reviewer’s verdict back to the same task. The plan, the rows, and the task-local files all survive; a new task would lose them. → Follow-up tasks for continue → Human in the loop for interleaving review and agent turns Decision: continue when the reviewer is refining the same goal. Start fresh only when the goal itself changed.
Restate the goal rather than forwarding the reviewer’s words. The agent sees one message, not the review thread the verdict came from, so “drop the bad ones” carries nothing.

​
5. Repeat

The task_id stays valid indefinitely. Steps 3 and 4 loop as many times as review needs, and the task keeps everything it learned.

Quickstart

The single-pass version of this.

Attach a CSV

The same loop, starting from rows you already have.