Read the logs
Logs are the conversations users had with a Web Agent, turn by turn, and the errors those conversations produced. Read them to see why the agent answered what it answered: the actions it ran, their input and output, the model and the version.
Find a conversation
To open the logs of an agent:
- In the portal, go to Web Agents and select your agent.
- Select the Monitoring tab.
The four figures at the start are Conversations, Errors, Error rate and Avg latency for the time range selected in Time.
Conversations lists the conversations, most recent first. Each row shows the conversation's title, the user and, when a turn failed, the error code. To narrow the list:
| Control | What it filters |
|---|---|
| Search | The conversation's title, the user's email or the visitor ID. |
| Status | Successful or Errors only. |
| Error code | Rate limited, Tool error, Timeout, Auth error or Context overflow. |
| Time | Last hour, Last 6 hours, Last 24 hours, Last 7 days or Any time. |
Who the user is
The row shows the first of these that the conversation has: the user's email, name, your ID for the user, or Anonymous with the visitor ID of the browser.
| Badge | Meaning |
|---|---|
| Verified | The identity came in a token signed by your backend. |
| Unverified | Your app passed the values to identify without a token. Anyone can send any value. |
| No badge | The app did not identify the user. |
Read a conversation
To open the transcript, select the row. The header shows:
- Success, or the error.
- The model and the version of the agent that answered. Version 5 is shown as
SDK v5. - The time the agent spent answering, added over the conversation, and the tokens used.
- The conversation's ID, with a control to copy it.
The transcript shows each message. Where the agent ran something, such as an action, the transcript shows a row with its name. To see what the agent sent and what came back, select the row and read Input and Output. A step that failed is marked error. Long values are cut and say how many characters are missing.
A microphone mark on a conversation means it has spoken turns.
To have the Advanced editor look into a conversation, select Audit in an interactive session. The editor opens with that conversation, checks it for failed actions and proposes a fix.
Read the errors
Errors lists every failed turn with its message, the conversation, the visitor, the model and the HTTP status. A failure appears here even when its turn is not in the transcript. To open the conversation an error belongs to, select View conversation.
Troubleshooting
| Symptom | Cause | Fix |
|---|---|---|
| A conversation is missing | A filter excludes it. | Set Time to Any time and Status to All statuses. |
The user shows as Anonymous and your app calls identify |
Enforce verification is on and the call sends no token, or the token did not verify. | See Enforce verified identity. |
| A conversation ran on an older version than the one you published | The page loads an environment that serves that version. | Deploy the version to that environment. |
Next steps
- Activity: what Faustine did with the agent, as opposed to what users did.
- Generate and run a test suite: check the actions before users reach them.
- Monitoring: logs and errors across every agent of the workspace.
Last updated November 13, 2026