Advanced sandbox: real contact, real mode, trace, regenerate and restart

Get the most out of the agent tester: inject a real contact, run actions for real with real mode, inspect the LLM trace, and regenerate or restart the conversation.

Jun 26, 2026

Watch this step in the video · The tester: trace, real mode and HTTP actions

The Sandbox (Live tester) is more than a test chat: it lets you test the agent with real data, run actions for real, and see exactly what the model decided. This guide covers the advanced features. If you're just starting, read How to test your agent in the sandbox first.

How to get there

  1. Go to Settings → AI Agents and open the agent.
  2. Go to the Sandbox tab.
  3. At the top you'll see three sub-tabs: Chat, Contact and Trace.

If you have unsaved changes, save before testing: the sandbox uses the saved version of the agent.

Real contact (Contact tab)

By default the agent talks to a synthetic contact. In the Contact tab you can give it real data so it personalises the reply:

  • Load a real contact: search by name, email or phone and pick it. The agent will resolve variables like name, company or tags from that data.
  • Or enter data manually: fill in name, email, phone, company, note and tags by hand.

This data is injected as context on every test message.

Real mode

The Real mode toggle (in the Chat header) changes what happens when the agent uses a tool:

  • Off (default): actions are simulated. The agent "says" it assigned or closed, but touches no data.
  • On: native tools (assign, close, add tag, etc.) actually run against the selected contact.

To turn real mode on you need a real contact loaded (without one, the toggle is disabled). The first time, it asks you to confirm with a warning: "actions modify workspace data". You'll see a red banner while it's active. Use it only when you want to audit the full end-to-end flow.

Trace

The Trace tab shows the detail of each turn: the LLM input and output and the tool calls it made. When the latest reply used tools, you'll see an amber dot on the tab to flag it without leaving Chat. Each tool call is labelled by how it ran:

  • Executed (real mode) — actually ran,
  • Simulated — touched no data,
  • HTTP action executed — your custom action (always real),
  • Error — something failed.

Use Copy JSON to grab the full trace for debugging or sharing.

Regenerate and restart

  • Regenerate: re-generates the last agent reply from the same input (handy when it came out weak and you want another variant without retyping).
  • Restart conversation: clears the test chat and starts fresh.
  • Copy conversation (JSON): exports the whole test dialogue.

Remember: the sandbox uses AI tokens

Sandbox tests are real model calls and consume your account's AI tokens (from your plan's monthly allowance or, once it's used up, from your Omnifox balance). The agent's daily token budget no longer stops any agent, so testing won't leave a production agent unable to reply.

If it doesn't work

  • I can't enable real mode: load a real contact first in the Contact tab. If you remove the contact while real mode is on, it falls back to simulation automatically.
  • The trace is empty: that reply didn't use tools. Try a message that triggers an action.
  • I'm testing the old version: save your changes; the sandbox uses the saved version.

Tips

  • Before going live with an agent, do a pass in real mode with a test contact to confirm it assigns, tags and closes as expected.
  • When a reply surprises you, open the Trace before touching the prompt: often the issue is a tool call, not the text.

Related articles

Was this helpful?

Related articles

Message us on WhatsApp