Agentforce enforces a 120-second (2-minute) maximum response time
per agent turn. If the agent doesn't complete its reasoning, tool calls, and response generation within that window, the session times out the error message leaving no trace in the STDM.
What we've checked:
- No timeout text in ErrorMessageText or OutputValueText
- AiAgentSessionEndType is NOT_SET for every session, so it's not usable
- Step durations show no spike near 120s; the killed step appears not to be recorded
Current approach: infer timeouts per interaction as "user Input with no agent Output." Complication: Output messages outnumber Inputs ~2:1, so welcome or progress messages may hide real timeouts.
Questions:
- Is there any STDM field or event that records a turn-level timeout?
- How are others separating timeouts from user abandonment?
- Is a timeout signal on the roadmap for Session Tracing?
#Data360
This appears to be a current observability gap in STDM. I’m not aware of any dedicated field that explicitly identifies a turn-level timeout.
Your approach of checking User Input with no final Agent Output is reasonable, but message counts alone may not be reliable because welcome or progress messages can create additional outputs.
A better approach would be to:
- Track each interaction/turn separately.
- Confirm that agent execution started.
- Check whether a final agent response was generated.
- If execution stops without completion, classify it as a suspected timeout.
- Where possible, validate it using API or channel-side logs.
I would also keep timeouts and user abandonment as separate categories, since STDM alone may not always provide enough information to distinguish them reliably.
At this time, I’m not aware of any publicly confirmed Salesforce roadmap item for a dedicated timeout signal in Session Tracing.
Thank You,
Venkat Yadav.
