Live Train

​

What Live Train is

The Train tab on each agent (AI Agents > [Your Agent] > Train) is a sandboxed environment for practicing with your agent. It runs the real agent — same instructions, same skills, same knowledge — inside an embedded preview widget you can iterate against without touching production. Use it to:

• Try edge-case questions and see how the AI responds.
• Test new instructions before saving.
• Verify that recently added knowledge sources actually surface.
• Practice escalation flows.
• Capture corrections and feedback that improve the agent over time.

​

How it works

The Train panel embeds an iframe of /preview/widget?...&mode=minimal running your real agent. Sessions in this preview are flagged so they’re excluded from production analytics and any AI training pipelines — your test conversations don’t show up in your Inbox or count toward usage stats. While you chat in the preview, the surrounding panel shows:

• Each message the AI generates with thumbs up / thumbs down ratings
• A correction box where you can write what the AI should have said, plus an Inbox Training button to save that correction as org-wide training (PII is scrubbed before it is stored)
• The conversation outcome when the session ends (resolved, escalated, dropped, etc.)

Live Train is proactive: you practice in a sandbox. Inbox training from tickets or the live feed is reactive: you capture lessons from real customer traffic. Both can write durable Inbox Training rows; the Train tab path is always org-wide and scrubbed. Your feedback is saved to a training log you can review later.

​

A typical Live Train session

1. Open your agent and click the Train tab.
2. Type a question your customers commonly ask: “How do I export my data?”
3. Read the AI’s response. Three things to check:
   • Did it search the right knowledge? Look for citations in the response.
   • Did it reference the user/company context? If not, you may need to send better context from the SDK.
   • Did it stay in scope? Compare against your instructions and Boundaries.
4. If the answer is good, give it a thumbs up.
5. If the answer is wrong or incomplete, thumb it down and write the better response in the correction box.
6. Move on to the next question.

After a few sessions, look at patterns in your corrections — they usually point to one of three things:

​

Iterating on instructions

Because Train uses the live agent, you can iterate without redeploying anything:

1. Edit instructions on the Configure tab.
2. Save the agent.
3. Switch back to the Train tab.
4. Ask the same questions again — the new instructions take effect immediately.

For changes to widget appearance, channels, or skills, do the same — save and re-test in the Train panel.

​

Practicing escalation

To practice escalation flows without filling your inbox with test tickets:

1. Set the agent’s escalation mode to conditional with a high confidence threshold (e.g. 70).
2. In Train, ask a question the AI clearly can’t answer (e.g. “What’s the status of my last order from Q3 2024?”).
3. The AI should escalate after a few attempts.
4. Note: the preview session skips actual ticket creation in the backend, so no real ticket is filed. The escalation flow is simulated so you can see how the AI behaves without filling your inbox.

​

When to leave Train and ship

Once Train conversations consistently meet your quality bar — knowledge cites correctly, tone matches your brand, escalation triggers at the right time — the agent is ready for real users. The Train tab stays useful even after launch:

• Use it to QA new knowledge after a doc update.
• Onboard new teammates by having them try the agent themselves.
• Reproduce specific customer issues to debug the AI’s reasoning.

​

Where to go next