Playground, publishing and rollback

Editing an agent never changes what your website visitors see. You work on a draft, try it in the playground, and then publish it as a numbered, immutable version. If a change turns out badly, you roll back.

The playground#

The playground is a chat window inside the agent builder that runs your draft configuration, including unpublished changes and agents that have never been published. Use it to:

• Ask your real top questions and check the answers.
• Try to break the agent: ask for off-topic things, claim to be an employee, ask for discounts, paste odd characters. See the prompt writing guide.
• Test tools. Custom HTTP tools really call your endpoint, so point them at a test environment if they create or change anything.
• Try different visitor attributes (name, email, custom values) and page URLs via the "test as visitor" settings to see how the variables fill in.
• Compare models by switching the chain and asking the same questions.

Playground conversations are marked as such: they do not count against your plan's monthly conversations, and are hidden from the inbox by default. They do not create real leads, send lead or handoff emails, or fire webhooks, so check those on the live widget after publishing. They still use your AI key, so they cost a little at your provider.

Publishing#

1. Open the agent and click Publish.
2. The draft is validated. If there are problems, publishing is blocked and the messages tell you what to fix.
3. Add an optional note (for example "Added returns policy, stricter pricing rules"). Notes make the version list readable later.
4. Confirm. The draft is copied into a new version with the next number and becomes the live configuration. New conversations use it immediately.

What validation checks#

• At least one step in the model chain, and every step has a model selected and uses a key that still exists.
• The chain is not longer than your plan allows.
• Custom HTTP tools: valid unique names (2 to 40 characters, lowercase letters, digits and underscores, not the name of a built-in tool), a description, and an https:// URL.
• The number of enabled custom tools is within your plan.
• Escalation rules that use live takeover are only accepted on plans that include takeover.
• Post-chat analysis is only accepted on plans that include it.

Validation cannot check whether your prompt is any good, whether your tool endpoints work, or whether your key has credit. Test those yourself.

What is and is not versioned#

A version captures the agent's configuration: prompt, knowledge, persona, models, tool definitions, lead and escalation settings, hours and limits. It does not capture:

• Agent secrets (the values behind {{secret.NAME}} in tools). They are stored encrypted and separately, so rolling back does not restore an old secret value.
• The contents of your AI keys, which live in AI providers.
• Widget appearance, which belongs to the site rather than the agent (see Widget customization).

Versions and history#

Each publish creates a version with a number, the author, the time and your note. Versions are immutable. Your plan determines how many are kept (Free 5, Starter 20, Pro and Agency 100 at the time of writing); when the limit is exceeded, the oldest are pruned, so do not rely on very old versions being available for rollback.

Each conversation records which version answered it, so when you see a bad conversation you can find out exactly which configuration produced it.

Rolling back#

1. Open the agent's Versions list.
2. Pick the earlier version you trust and choose Roll back (restore).
3. Confirm. The old configuration is published again as a new version at the top of the list, so history stays complete.

A rollback can fail validation if the old version relies on something that has since changed, such as a deleted key or a feature your plan no longer includes. You will see which, and can fix it in the draft before publishing.

Disabling an agent#

An agent has an enabled switch. A disabled agent is not used to answer. If the site has no enabled, published agent, the widget reports itself as unavailable and stays hidden. This is the quickest way to pause chat in an emergency without editing your website. If a site has several agents, its default agent answers; otherwise the first enabled, published agent does.

A safe workflow#

1. Change one thing in the draft.
2. Replay your test questions in the playground.
3. Publish with a descriptive note.
4. Chat with the live widget once as a sanity check.
5. Watch Conversations over the next day for surprises, and roll back if needed.