Skip to main content

Logs and traces

Logs reveal what happened behind a conversation, so you can find the root cause when an agent behaves unexpectedly.

Open logs and traces​

Logs show what happened behind a conversation: which tools the agent called and exactly what was sent to and returned by the language model. Use them to find the root cause when an agent behaves unexpectedly.

  1. Select a session and click Logs under the chat.
  2. The Logs drawer opens with two tabs: Trace Logs and System Logs.

In the drawer you can:

  • Search — type in the search field to highlight matches and step through them with the arrow buttons.
  • Copy to clipboard — copy the full text of the active tab.
  • Refresh — reload the logs, useful while a session is still running.
note

Trace Logs are available on paid plans. On the Free plan, the drawer shows System Logs only.

System Logs​

System logs record the session lifecycle: session opening and initialization, saving and closing, tool and skill calls, and other events that affect agent behavior. Use them to confirm what the agent did — for example, whether a handover or notification tool was actually called.

Trace Logs​

Trace logs capture every LLM call in the session, separated by ####### markers. Each call has three parts:

  1. Agent name — which internal agent handled the request. Several agents can take part in one reply (for example, the main conversation agent plus a form agent that records results).
  2. Prompt — the full prompt sent to the model: the main task, identity and workflow context, chat history, available tools, and system info such as the current date.
  3. Generation — the model's response, structured as an Action (what the agent decided to do) and an Action Input (the parameters for that action).
tip

To debug one specific reply, search the trace for a distinctive phrase from that reply, then read the surrounding prompt and generation.

A typical investigation: read the chat to understand what the customer experienced, check session results for missing or wrong values, open System Logs to confirm which tools ran, then open Trace Logs to see whether the prompt contained the right context and how the model responded.

Was this article helpful?