I fixed an AI agent's mistake. It kept undoing my fix.
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
4 weeks ago
I fixed an AI agent's mistake. It kept undoing my fix. In a shared browser, the AI couldn't distinguish my edits from its own errors, repeatedly reverting changes and wasting tokens. So, I built ActionTrace, which tracks human vs AI browser actions. Now instead of guessing, logs show:
[HUMAN] edit vs [AI] edit
GitHub: https://github.com/nvaishnavisriyaa/ActionTrace
Demo: https://youtu.be/WKw3Eqf8n_E
Question for Otto team: When humans intervene in ITSM/CSM agent workflows, how does Otto distinguish intentional human changes from AI errors? Is this context available to the agent in real-time or only in audit logs?
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
2 weeks ago
This is an interesting approach. I think the distinction between human actions and AI-driven changes is especially important when multiple users and agents work on the same ITSM records.
ServiceNow does provide some visibility into this through the AI Workflows panel.
Agent-modified fields display an indicator identifying AI-driven changes.
You can review agent execution history through the AI Workflows panel.
Check the sn_aia_execution_plan table for execution details.
Review execution tasks, messages and tool executions to understand what the agent changed.
However, the important question is whether human edits are captured as real-time context and passed back to the agent before its next action. Having an audit trail does not necessarily mean the agent can use that information to avoid overwriting human changes.
I'd be interested to know whether the Otto team has a built-in mechanism for this or whether customers need to implement their own action-tracking logic.
Useful references:
