# Test, publish and versions

Run a draft agent in the sandbox, read the pre-flight and publish dialog, understand where a publish lands and how versions, rollback and promotion work.

A draft is private until you publish it. The **Test** tab runs it live in the sandbox, the **Deploy** tab shows where it will run and what is still missing, and **Publish** checks the draft before it becomes a version. Every publish creates a version; the agent's page is where versions are compared, rolled back and promoted.

**Who can use it:** the **Builder** role or higher. Promoting to production and rolling back need release authority.

Video: Test, publish and versions (2:21), https://katonicdocsmedia.blob.core.windows.net/docs/test-publish-and-versions/9afd8aeb9047/index.html

  
**Step 1. The draft**

Our Expense Policy Assistant is configured: five of eight sections, three knowledge bases, one skill. Now we open **Test**.

  
**Step 2. Test**

The **Test** tab runs the draft through the same executor that serves published agents, without publishing. Type a question, and run it. The answer streams in, with its sources. The **Run trace** on the right shows the real steps: run started, configure, the retrieval tool, compose. Duration and tokens sit above them.

  
**Step 3. Deploy**

**Deploy** shows where the draft will run. **Environments** lists your tenant's lanes, none configured here yet. **Surface** is how people consume it: Chat, Panel, Full app or Voice. **Pre-flight** on the right lists what is still missing, and **This version** sums up the draft.

  
**Step 4. Publish**

**Publish** checks the draft first. The subtitle says it is ready, with two warnings to review. Under **Who can use it**, Organization is selected. Under **Where it runs**, we add the **REST API**. **Section status** marks every section Required, Review or Optional, and whether it is done. A blocker would name what is missing, with **Go to section** beside it. Two warnings, not blocking: no memory configured, and the tenant's environments could not be read. The button reads **Publish**. **Published, not live yet.** Production and the Work Room are not serving it, because it may be waiting on approval. **View agent** opens its page.

  
**Step 5. Versions**

Every publish adds a row to the agent page's **Versions** tab: draft, live or initial, with who made it and what changed. **Diff** compares a version with the one before. **Roll back** makes an earlier version live again; picking a newer one files a promotion request instead.

## Test the draft

![The Test tab of the wizard: the Test prompts list with three saved prompts and the Batch run summary on the left, the sandbox conversation pane with Run in the middle and the message box, and the Run trace panel with Duration, Tokens and Steps on the right.](https://docs.katonic.ai/img/agents/test-publish-and-versions/wizard-test.png)
*Test runs the draft through the same executor that serves published agents, without publishing. Save prompts to rerun them, and read the trace to see each step of the run.*

The **Test** tab runs the draft through the same executor that serves published agents, with your prompt, tools and knowledge, and without publishing. Type a message in the box or pick a saved prompt and select **Run**; the answer streams in and the **Run trace** on the right shows the duration, tokens and steps of the real run: configuration, tool calls and the model call.

**Test prompts** on the left keeps prompts you want to rerun; add one with the plus button. **Batch run** counts what ran this session, how many succeeded or failed and the average latency, and **Run all** reruns every saved prompt. The chips under the message box show the environment and how many knowledge sources, tools and skills the draft has. A draft with no system prompt has nothing to run; add one in Behavior first.

## Read the Deploy tab

![The Deploy tab: Ship this draft with the Environments list, Development and Production, the Surface choices Chat, Panel, Full app, Voice and Live, the API endpoint with a curl example and the start of the Channels card, with the Pre-flight and This version panels and Publish on the right.](https://docs.katonic.ai/img/agents/test-publish-and-versions/wizard-deploy.png)
*Deploy shows where the agent will run: its environments, surface, channels and rollout. Pre-flight lists what is still missing before you publish.*

**Ship this draft** lists the **Environments** your tenant has, or says none are configured yet. **Surface** is how people consume the agent (**Chat**, **Panel**, **Full app**, **Voice**; **Live** is coming soon) and can be set once the agent is published. **API endpoint** shows the address callers will use with a bearer token. **Channels** shows that the Work Room is on and the API can be switched on; Slack is coming soon. Under **Production rollout**, **Instant** is available; **Staged** and **Shadow** are coming soon.

**Pre-flight** on the right lists what is still missing: a system prompt, a guardrail profile, tools, knowledge, and a run of the Test tab. **This version** sums up the draft. If your organisation publishes into a lane and has no production lane, a notice above the environments says promotion is unavailable until an administrator creates one.

## Publish

![The Publish agent dialog: Reachability with Team or Organization under Who can use it and Work Room or REST API under Where it runs, the Section status list marking Identity and Behavior as required, and the start of the blockers list with a Go to section link on each.](https://docs.katonic.ai/img/agents/test-publish-and-versions/publish-modal.png)
*Publish checks the draft first. Blockers must be fixed and each has a Go to section link; warnings are for review. Reachability decides who can use the agent and where.*

Select **Publish** in the build header. The subtitle says how many blockers remain, or that the draft is ready with warnings to review, or that all checks passed. Under **Who can use it**, choose **Team** or **Organization**; under **Where it runs**, tick **Work Room**, **REST API** or both. **Section status** marks each section **Required**, **Review** or **Optional** and whether it is **Done**; a blocker names what is missing, such as a display name, a slug, an owner team, or a system prompt of at least 30 characters, with **Go to section** beside it. Warnings do not block. When the draft carries settings the runtime treats as advisory, **What's enforced** lists each as **Enforced**, **Advisory** or **Not enforced (not in edition)**, and you tick an acknowledgement before the button enables. The button reads **Publish to production**, or **Publish to** the lane your tenant publishes into, and **Publishing…** while it runs.

## What the success screen means

The screen after a publish tells you where the version landed:

- **is now live**: the new version serves traffic. **Open in Work Room** tries it, **View agent** opens its page, and the REST endpoint can be copied.
- **published to** a lane, **promote when ready**: the version is deployed to that environment only. Production still serves the previous version until you promote from the agent's page.
- **not live yet**: the version is published but production and the Work Room are not serving it; it may be waiting on approval. Open the agent page to see its status.
- A warning that model availability was not verified means the AI Gateway could not be reached to confirm a model; check the gateway's models in the Control Room.

## Versions, rollback and promotion

Each publish adds a row to the agent page's **Versions** tab, tagged **draft**, **live**, **initial** or **not live**. **Diff** compares a version with the one before it. **Roll back** makes an earlier version the one production serves again. Picking a version newer than the live one files a promotion request instead, and production keeps serving until a release manager approves it. Attaching or detaching knowledge from the agent's page also publishes a new version. Details are in [The agent's page](/agents/agent-page).

## Troubleshooting

- **This draft has no system prompt yet, so there is nothing to run.** Add one in Behavior, or name the agent in Identity, then run again.
- **Couldn't read this tenant's environments.** The publish destination is not confirmed; check the Control Room's Environments page before publishing.
- **Fix N blockers** stays on the button until every blocker is resolved.

## Related

- [Build an answer agent](/agents/build-an-answer-agent)
- [The agent's page](/agents/agent-page)
- Environments
- [Using agents in the Work Room](/user-guide/work-room/using-agents)

Source: https://docs.katonic.ai/agents/test-publish-and-versions
