# Read the results and apply the winner

Read a completed run on Results, compare the winning prompt with the baseline evaluator by evaluator, apply it from Promote and roll it back if needed.

A completed run ends on **Results**, which shows what the winner is and how much better it scored. **Promote** is where that prompt is published as a new version of the agent, first to the sandbox lane. Both tabs name the winner **cand-n** after the trial that produced it.

**Who can use it:** the **Builder** role or higher. Applying writes a new version of the agent and publishes it, so it takes the same right as publishing the agent.

## Results

![The Results tab: a winner card titled Optimized prompt with a candidate badge, the lift percentage, trial count, final score and token change, buttons Copy as draft, Promote and Discard, a row of delta cells for Objective, Relevancy, Helpfulness and Safety, then a two-column diff of the baseline and optimised system prompts.](https://docs.katonic.ai/img/agents/optimizer-results-and-promote/optimizer-results.png)
*Results: the winning candidate with its lift over the baseline, each evaluator's baseline and optimised score, and the baseline and optimised prompts side by side with the changes marked.*

Until the run completes the tab reads **Results populate when the run completes.** with the current status and the trials done. A failed run shows **Run failed** with the reason and a **Technical details** disclosure instead.

The winner card is titled **Optimized prompt** with the candidate badge, then the lift as a percentage over the baseline, how many trials ran, the final score, and the size of the change as tokens added and removed. A run whose baseline could not be scored reads **baseline not scored** here and claims no lift. Once applied, the card adds **applied to live prompt**.

- **Promote** moves to the Promote tab.
- **Copy as draft** and, in the diff's footer, **Copy optimized** open Prompt Studio. The optimised text is not carried into a new draft in this release; select it in the **Optimized** column to copy it.
- **Discard** is disabled in this release. A run you do not want is deleted from History instead.

Below the card, one cell per objective: **Objective** shows the baseline score, an arrow, the optimised score and the percentage change; each evaluator shows its baseline and optimised scores with its weight.

### The diff

**Base** is the live system prompt and **Optimized** the winner. Each header carries its token count and the number of lines removed or added, and the body marks the changes line by line.

### Per-evaluator breakdown

One row per evaluator with its weight, the **Baseline** score, the **Optimized** score, the change in points and a **Distribution** bar. **By eval** is the view available; **By slice** and **By case** are disabled. When a run returns no per-evaluator scores, the table says so and shows the headline objective only.

## Promote

![The Promote tab: a hero reading Optimized prompt cand-5 ready to apply with the scores against the baseline and the Customer Support Agent it applies to, a Lift vs baseline stamp, a What you're shipping card listing the words added, 01 Target deployments naming the agent with a current to optimized transition at 100% binary apply and Open agent deployment view, 02 Changelog, and the sticky bar saying the apply publishes a new version to the sandbox lane, with Cancel and Publish to sandbox now.](https://docs.katonic.ai/img/agents/optimizer-results-and-promote/optimizer-promote.png)
*Promote: the winner ready to apply, what you are shipping as word-level additions and removals, the target agent, the changelog, and the sticky bar whose Publish to sandbox now publishes a new version of the agent to the sandbox lane.*

**Promote** needs a completed run with an optimised prompt; before that it reads **Run not complete**. A failed or cancelled run reads **Run failed** or **Run cancelled** and says there is nothing to promote. A run that produced no prompt change reads **This run produced no optimized prompt - nothing to apply**, and a run whose best candidate did not score above the baseline reads **No improvement: keep the current version**, with the apply turned off.

The hero names the candidate **ready to apply**, the score it reached on the trial split against the baseline, and the agent it will be applied to, with a **Lift vs baseline** stamp. **What you're shipping** sums up the change as words added and words replaced or removed across the system prompt, plus a **Score-based selection** line with the two scores; a candidate with no textual change is flagged as a no-op.

  
**Step 1. Check the target**

**Target deployments** names the agent, labelled **Live agent prompt**, with its id and tenant, and the transition **current → optimized** at **100%**, a binary apply. Promoting the new version to other environments and reviewing the version history happen on the agent's page: **Open agent deployment view →** takes you there.

  
**Step 2. Read the changelog**

**Changelog** shows what will be recorded: the apply, the score change on the trial split, the diff size, the run id, the aggregation and the number of evaluators.

There is no canary plan and there are no rollback rules. Both used to sit here as disabled controls, a 5 to 100 percent ramp with stage holds and four live-metric guard rails; nothing wired either, so they were removed rather than left promising a safety net that did not exist. The one operation here publishes a new version of the agent, and rolling back means re-publishing the previous version, which the applied banner does.

  
**Step 3. Apply**

The sticky bar reads the apply back: the agent, the words added and removed, the lift, and that it publishes a new version to the sandbox lane while production keeps its current version until you promote. **Publish to sandbox now** asks **Publish the optimized prompt of** the agent **to sandbox now?** and, on **Publish to sandbox**, creates a new version of the agent with the optimised prompt and its tuned parameters and publishes it to the sandbox lane. Production keeps serving its current version until you promote the new one from the agent's page. **Cancel** returns to Results without applying.

![The Promote tab after publishing: an Applied banner reading The optimized prompt is now live on agent Customer Support Agent with a Roll back to previous version button and Next: Compare versions in the Profiler, and the sticky bar reading Already applied to agent Customer Support Agent, saying the optimized prompt is published to the sandbox lane, with Back to results and a disabled Applied button.](https://docs.katonic.ai/img/agents/optimizer-results-and-promote/optimizer-promote-applied.png)
*After Publish to sandbox now: the Applied banner offers Roll back to previous version, and the sticky bar says the new version is in the sandbox lane until you promote it from the agent page.*

## Verify

After publishing, an **Applied** banner reads **The optimized prompt is now live on agent** with the agent's name, and offers **Compare versions in the Profiler** as the next step. The bar reads **Already applied to agent** and says the optimised prompt is published to the sandbox lane, and the button reads **Applied**. The agent's versions show the new version in the sandbox lane, and the run reads **applied** in History. To put it in production, promote it from the agent's page.

## Roll back

**Roll back to previous version** in the Applied banner asks **Roll back this agent to its previous published version?** and, on **Roll back**, re-publishes the version that was live before the apply. The banner then reads **Rolled back**.

## Troubleshooting

- **Apply failed.** The reason follows in the banner. The run must be completed, the agent still published and you must hold the right to publish it.
- **baseline not scored.** The baseline evaluation did not run, so no lift can be stated. The prompt can still be applied on its absolute score.

## Related

- [Follow a run and inspect its trials](/agents/optimizer-follow-a-run)
- [Agent page](/agents/agent-page)
- [Prompt Studio](/agents/prompt-studio)
- Agents in the Control Room

Source: https://docs.katonic.ai/agents/optimizer-results-and-promote
