> ## Documentation Index
> Fetch the complete documentation index at: https://docs.getoneprofile.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Train an agent

> Rate the answers you disagree with, then work through a training round and approve the changes you want.

Rate the answers you disagree with, right from the cell that holds them. Then run a training round: Oneprofile rewrites the prompt to fix what you marked bad, and you approve the change before it ships.

<Frame caption="Reviewing one row during a training round.">
  <img src="https://mintlify.s3.us-west-1.amazonaws.com/oneprofile/images/agents/train-an-agent.png" alt="The training screen showing one row's output above Good and Bad buttons, with a note field open under Bad explaining what is wrong." />
</Frame>

## Rate a cell any time

Right-click any cell to mark it good or bad, or open its drawer and use the thumbs at the bottom. A note on a bad rating is optional there, but it is the detail a rewrite can act on.

<Steps>
  <Step title="Rate from the grid">
    Right-click a cell, or a selection of them, and choose **Mark as good** or mark it bad. Add a note if you want to.
  </Step>

  <Step title="Or rate from the drawer">
    Open the cell and click the thumbs at the bottom, then click **Add note** to say what was wrong.
  </Step>
</Steps>

## Start a training round

<Steps>
  <Step title="Open Train">
    From the agent editor, click **Train**.
  </Step>

  <Step title="Choose training data">
    Pick **Enter manually** to type example rows, or **Use a table** to run against real ones. Click **Start training** once your rows are ready.
  </Step>
</Steps>

## Review each row

On this screen, a bad verdict always needs a note. Good needs nothing and moves you straight to the next row.

<Steps>
  <Step title="Mark it">
    Click **Good**, or click **Bad** and describe what is wrong and what it should have returned instead.
  </Step>

  <Step title="Reach the summary">
    Once every row has a verdict, the summary counts how many were good and how many were bad.
  </Step>
</Steps>

## Approve or reject the revision

With at least one row marked bad, click **Train & review** to have Oneprofile propose a rewritten prompt. Compare it against the current one before it becomes the agent's.

<Steps>
  <Step title="Read the proposal">
    Compare the current prompt against the proposed one, line by line.
  </Step>

  <Step title="Decide">
    Click **Accept & re-run** to save it and re-run your rows. Or click **Ask for another version**, or **Reject** to discard the proposal.
  </Step>
</Steps>

<Warning>
  Accepting a proposal saves a new version and re-runs your training rows. Every table column using this agent picks up the change on its next run.
</Warning>

## Good to know

<AccordionGroup>
  <Accordion title="What happens if I discard a training round?">
    The rows and feedback are kept as history, but the round ends and the agent is free for a new one. A revision you already approved stays applied.
  </Accordion>

  <Accordion title="Does rejecting a proposal lose my earlier rounds' feedback?">
    No. Rejecting discards only this round's proposal and its notes. Feedback from earlier rounds is kept.
  </Accordion>

  <Accordion title="What if the new prompt still gets a row wrong?">
    Mark it bad again with a fresh note. A row left unchanged since your last note is flagged, so you know it still needs attention.
  </Accordion>

  <Accordion title="Can I change the effort level mid round?">
    Yes. Open **Change effort** from the training screen's menu, without leaving the session. The next round uses whatever you save.
  </Accordion>
</AccordionGroup>

## Next steps

<CardGroup cols={2}>
  <Card title="Agent versions" icon="history" href="/agents/agent-versions">
    See what an approved revision does to earlier versions.
  </Card>

  <Card title="Read an agent run" icon="list-tree" href="/agents/read-an-agent-run">
    Look at the trail behind an answer before you decide it is wrong.
  </Card>
</CardGroup>