FAQ

These questions cover the first build, controlled launch, and daily operation.

Creating the first Agent

Should I use a Template or start blank?

Use a Template when its job and boundaries resemble yours. Review every generated field before creating the Agent.

Start blank when no Template fits or when you already know the exact behavior you need. Both paths return to the same Agent Editor and verification flow.

Do I need a complete Knowledge Base before testing?

No. Start with the smallest source of truth required for the real situation.

Add more content when a failed case shows that missing or outdated knowledge is the cause. Uploading every available document can make scope and diagnosis harder.

How many cases do I need?

There is no universal number.

Begin with one real situation and a few nearby boundaries. Before a controlled pilot, expand the set according to frequency, consequence, tool use, and handoff risk. Formal acceptance may require a locked version, case set, Standards, and must-pass conditions.

What makes a strong Standard in Test Suite?

Write observable requirements, not a vague hope or one exact sentence.

State what the result must do, what it must avoid, and when the Agent must hand off. Another reviewer should be able to decide pass or fail without guessing.

Publishing and release

Do I need to publish every time I change the Agent?

Draft changes do not affect users. Publish the intended Agent version when the change is ready.

If Channel settings also changed, publish the Channel separately.

Why does the live URL open, but there is no usable Agent?

The most common reason is that the Channel exists, but the Agent version has not been published.

Return to AI agents, publish the intended version, and test the live URL again.

When should I use a Web Channel instead of API access?

Use a Web Channel for the fastest path to a hosted pilot or customer experience.

Use API access when your own product should control the client and integration behavior.

Conversations and improvement

How do I reply when a user is waiting for a person?

Open the conversation marked Needs human reply.

If human reply is off, use Open CRM and start the handoff for that customer-Agent pair. Write the reply or generate an AI draft, review the final content, and send it through the original Channel when available.

Does AI draft send a reply automatically?

No. AI draft only creates a proposal.

A workspace member must apply or copy it, edit the response, review any proposed form, and click Send.

How do I turn a bad conversation into a reusable case?

Open the thread in Conversations and use Add Case.

Keep the necessary context, write a checkable Standard, diagnose the cause, and rerun the case after the focused change.

Does one passing case prove the Agent is ready?

No. It proves only that the checked version met the case conditions in that run.

Use must-pass, boundary, tool, handoff, and regression cases in proportion to the release decision. Continue reviewing real conversations after launch.

Data and access

Why is an updated Notion or Google Drive file not affecting the Agent yet?

Treat connected material as a managed snapshot.

Refresh the connected copy after a meaningful source change, then rerun the relevant cases.

When should I use Whitelist mode?

Use Whitelist when the rollout should stay controlled, such as an early pilot or limited expert review.

Manage the allowed audience in Customers.

What is the difference between Organization and Workspace?

Organization is the broader team boundary.

Workspace is where a specific set of Agents, Channels, Conversations, Customers, operators, and data boundaries are managed together.

More help

For a repeatable problem in a real conversation:

Handle the user's immediate need.
Use Add Case to preserve the situation.
Decide whether the issue is the expected behavior, instructions, source, tool, or scope.
Make the smallest relevant change.
Rerun affected cases before publishing again.