Get-started FAQ

Why can I not create another agent?

Creating an agent requires the workspace's agent-create permission. The Create New Agent dialog also checks the active-agent limit. When the current plan reaches that limit, the dialog disables the name field and asks you to upgrade. You can also reduce the active-agent count. An agent scheduled for deletion does not count toward the limit. Existing agents remain available from the Agents dashboard.

Where did my new agent go?

After you create an agent, AgentStack opens its Analytics page. Return to the Agents dashboard to see every active agent. If you select an agent card, AgentStack also opens Analytics. The card menu gives direct links to Settings and Analytics.

Which source do I add first?

Use the smallest authoritative source that answers a question you can verify. A single public policy page or a focused Q&A pair is easier to check than a broad website crawl. When files or Notion pages hold the real source of truth, add them from the Content page.

Why does the agent not know content that I just added?

Website ingestion is asynchronous. Under Sources → Content, wait for the website status to become Completed. Then open the Pages tab and confirm that AgentStack crawled and indexed the relevant page. A source that is still Crawling or Processing is not ready for a final quality test. If the status is Failed, read the error next to the source, then correct that error.

Why did my Playground improvement disappear?

Apply & Test applies the current model and system prompt to the Playground test chat. It does not save the configuration. When the editor differs from the stored agent, an Unsaved changes badge appears. After the candidate passes your tests, select Save to Agent. Reset restores the last saved values.

Why does the website widget behave differently from Playground?

Check four things:

1. You saved the successful Playground model and prompt to the agent.
2. Settings → General → Public is enabled. A non-public agent does not show its external widget.
3. Domain Protection allows the site's hostname, or you leave it empty to allow all domains.
4. The site uses the current embed code from Deploy.

Then reproduce the question in a private browser window. Look for the conversation in Activity. This test separates a cached-page problem from an agent-configuration problem.

Which model do I choose?

Begin with the default model, currently GPT-5.2. Grade it against a fixed acceptance set. Switch models when another model gives a repeatable improvement in accuracy, boundary behavior, or latency. Switch only when the improvement justifies its credit cost. The selector is the current source of truth for available models, recommended labels, and credits. See Choose a model.

Does model comparison cost more?

Each active comparison panel sends a separate model request. Playground shows the combined credit cost and active panel count before a synchronized send. Pause panels that you do not need. Compare models on hard, representative questions. Do not use credits on greetings.

Do product facts belong in the system prompt?

Usually not. Keep stable behavioral instructions—role, tone, boundaries, uncertainty, and escalation—in the system prompt. Put policies, prices, procedures, and other changing facts in sources. This split makes factual updates easier to own. It also reduces conflicts between instructions and knowledge.

What does an agent do when the answer is missing?

Define the expected behavior. When the available knowledge does not contain the answer, a safe first-agent prompt tells the agent to state its uncertainty. The prompt also gives a configured support contact or a manual next step for this case. An actual transfer requires a configured handoff capability. Before deployment, test the agent with at least one question that has no answer in your sources.

Do I need to configure actions before launching?

No. A first agent can answer from sources without external actions. Stabilize answer quality first. Add integrations and actions later, with narrow permissions and careful testing. Actions can change another system. Text output cannot.

What do I test before publishing?

At minimum, test several known facts, a paraphrase, an ambiguous request, an unsupported question, and a boundary case. Save the final configuration. Then repeat a supported and an unsupported question on the real customer channel. The response-quality guide gives a reusable checklist.

Where do I go next?

Follow Create your first agent for the complete setup sequence. Then use these guides for more setup:

• Sources for deeper knowledge configuration
• Web widget for installation
• Channels for other customer surfaces
• Inbox for human handoff