Conversation trace
The AI conversation trace shows every turn your AI agents took with one contact. A turn is one round: a message or sentence comes in, the agent thinks, and it replies.
Use it when an answer looks wrong and you want to know why. It shows what the agent read, which skills it used, and whether anything failed.
Find it: two ways.
- Sidebar → Inbox → Messages → open a conversation → ⋯ (Conversation menu) → AI trace.
- Open a contact's panel → ⋯ (More contact actions) → AI conversation trace.
Click Back to … at the top to return to where you came from.
When to use it
| Problem | What to look for in the trace |
|---|---|
| "The agent gave the wrong price." | Context retrieved: did it read the right knowledge? |
| "It said it booked, but nothing was booked." | The skill call: did it say "refused"? |
| "It went silent or answered strangely." | A red error pill or a Fallback event. |
| "It refused to send a message." | A Guardrail event with the reason. |
If you only need the words that were said, Agent activity is quicker.
Read the trace
Turns are grouped by conversation and day, like "WhatsApp conversation · 12 Oct 2026 · Asha · 6 turns".
Each row shows:
| Part | What it means |
|---|---|
| Turn N and the first words | Which turn it is and what was said |
| Time | How long ago. Hover for the exact time. |
| Channel and agent | Where it happened and which agent answered |
| Model | Which AI model handled the turn |
| Tokens | How much text went in and came out, like "1.2k in · 180 out" |
| Speed | How long the turn took |
| Tool calls | How many skills the agent used |
| N errors (red) / N warnings (amber) | Something failed, or a guardrail or fallback kicked in |
Scan for red pills first. Rows with errors open by themselves.
Open a turn
- Click a row. You'll see: the turn's steps, in order.
- Read the steps:
- Input received: what came in.
- Context retrieved: the knowledge or contact details the agent read.
- LLM call: what the agent sent to the AI model and what came back.
- A skill line, like "send_whatsapp_template called" or "… refused" with a reason.
- Guardrail: a safety rule that blocked something, with its code.
- Outbound sent: what was sent to the contact.
- Error or Fallback: what went wrong.
- Click Show … under a step to see the full details. Click the copy icon to copy them for support.
Filter and refresh
| Control | What it does |
|---|---|
| Search what was said | Finds turns by their words |
| Agent | Shows one agent's turns |
| Channel | Shows one channel |
| Period | Last 24 hours, Last 7 days or Last 30 days |
| Refresh (circle arrow) | "Look for new turns" without reloading the page |
| Load more | Loads older turns |
Common problems
"No AI turns yet". This contact hasn't had an AI conversation. Turns appear as soon as an agent answers.
"No turns match". Change or clear the search and filters.
I found the problem. How do I fix it? The trace is read-only. Fix the agent's instructions, knowledge or skills, then test with Try it live. See Fixing a misbehaving agent.