The “How this reply was made” panel shows how an AI agent produced a reply in Inbox: the information it received and the steps it took. Use it to understand a reply that seems incorrect, incomplete or unexpected. The panel is read-only, so you can’t change the agent’s configuration from it.
Before you start:
- “Adjust” only appears in workspaces that have the “Optimization: feature. See Review and improve AI agent responses with Optimization.
- Records are kept for 90 days, so you can only see how replies from the last 90 days were made.
When to use it
Execution trace can help you:
- Check what information the reply used, such as Knowledge, contact properties or handover context from another user.
- See which steps the agent took, and which failed, were skipped or were not reached.
- See which integrations the agent used, such as Shopify actions or custom APIs.
- See how many credits the reply used and whether it was a Basic or Advanced answer.
Open the panel
You can follow the steps below to access the Execution trace:
- Click on the
icon to go to Inbox - Open a conversation that contains a historical AI-generated message.
- Click “Optimize reply” under the AI message. If teammates have reported the message, the button also shows how many.
The “How this reply was made” panel opens on the “How it answered” tab.
AI Agent reply
“AI Agent reply” shows the reply that the panel describes.
If your workspace has an AI Wallet, a line under the reply shows how many credits the reply used and whether it was a Basic or Advanced answer. For how credits are charged, see “Managing your AI Wallet”.
⚠️ If the credit line is missing, the AI Wallet has no credit record for that reply. This doesn’t mean the reply was free.
Review what it received
“What it received” shows the information the agent had while it created the reply, and where each piece came from. It can include:
Item |
What it shows |
“Knowledge” |
A Knowledge source, marked “Used in reply” or “Not used in reply” when known. Click “View details” to see the passage the agent used, then “Open in Knowledge Base” to open the source. If the passage wasn’t recorded, the panel says so. |
“Knowledge search” |
Sources found during the knowledge search that the reply didn’t use. |
“Contact property” |
A contact property value the agent used. Very long values are shortened in the panel, but the agent used the full value. |
“Handover context” |
Context the agent received when the conversation was handed back to it, and who added it. See “Manage AI agent sessions in Inbox with Human-AI Conversation Control”. |
“Conversation” |
Messages exchanged while another user handled the conversation, before the agent resumed. Click “View details” to see them in order. |
If the reply didn’t use any information, this section says so.
Review what it did
“What it did” lists the steps the agent took to create the reply, in order, in plain words. If the agent used an integration outside a procedure step, the step names the integration.
If the agent answered directly without taking any steps, this section says so.
Understand step statuses
Steps that ran normally have no status. Other steps show one of these:
- “Failed”: the step didn’t finish.
- “Skipped”: the agent skipped the step.
- “Not reached”: the agent didn’t get to the step.
The panel doesn’t show steps as loading or in progress.
Report a reply
To report a problem with a reply, open the “Report” tab in the same panel. Your report is sent to Optimization Hub for review. For what happens next, see “Review and improve AI agent responses with Optimization”.