Skip to Content
How-to guidesHow-to guidesRead agent self-reports

Read agent self-reports

When an agent encounters an error and wants to escalate to its human owner, it calls report_self_diagnostic via the AgentValet MCP. Those reports land at /agent-reports.

Page layout

  • Title: Agent Reports
  • Subtitle: “Self-reported diagnostics submitted by your agents. Auto-refreshes every 30 seconds.”
  • Count badge: “{N} reports” in the header

Filters

  • From / To — date range
  • Agent — multi-select dropdown
  • Severity — chips for debug / info / warn / error / critical (toggle on/off)
  • Search — free text against message body

What each row shows

  • Time
  • Severity (badge)
  • Agent
  • Platform (if the report was platform-specific)
  • Code (the agent’s classification — e.g., upstream_401_after_refresh)
  • Message (the agent’s one-sentence narrative)

Click for the detail modal.

The detail modal

Header: AGENT REPORT (small) and the report’s code (large, or “diagnostic” if no code was supplied).

Body:

  • Time
  • Severity
  • Agent
  • Platform
  • Endpoint (the proxy path that triggered the error, if relevant)
  • Message
  • Correlation ID — the request_id you can use to find the originating audit entry. “Look up request_id = {id} in the audit log”
  • Forensics section (if available) — source IP and user agent of the calling agent
  • Context section — full JSON the agent included

How agents create reports

Three patterns:

  1. In response to a use_platform error that returned a report_hint block. The agent asks the user “Want me to lodge this with your AgentValet owner?” and on yes, calls report_self_diagnostic with the correlation_id from the hint.
  2. Proactively, on its own initiative — e.g., when noticing it’s failed the same call three times in a row.
  3. Programmatically — anything the integration thinks is worth a record.

Acting on reports

Reports are read-only on this surface — you can’t reply to the agent from here. What you can do:

  • Open the linked audit entry by request_id to see the actual failed request
  • Reach out to whoever’s running the agent (especially if it’s a teammate) to share what you saw
  • For credentials-expired flavours, reauth the platform
  • For repeated misbehaviour, revoke a scope or suspend the agent

Notifications

By default, error and critical severity reports send a push notification (if you’ve enabled push). Configure per-severity notifications under /settings → Agent report notifications.

Empty state

“No reports match your filters. Agents lodge reports via the report_self_diagnostic MCP tool, or in response to a use_platform error containing a report_hint.”

Next

Manage anomaliesProfile, password, passkeys