The loop that hit me hard:Web Agent types wrong value in field.
I manually correct it.
Web Agent continues: "Great, I typed it correctly" → moves to next step.
Everything after that fails because it thinks it made my edit. It had no idea I touched the browser at all. Built ActionTrace to fix this exact blind spot:
[AI] typed "ABC" → wrong
[HUMAN] typed "XYZ" → this is the real value now
Now the Web Agent knows that an error didn't happen magically, human made the change - without human needing to explicitly say it.
🔗 GitHub:
https://github.com/nvaishnavisriyaa/ActionTrace
🎥 45-sec demo:
https://youtu.be/WKw3Eqf8n_E
Question: When human edits a field mid-flow in Service Cloud, does Agentforce get live context that "human made this change"? Or does it assume it made the change itself? Anyone seen this?
#Agentforce #AI #AgenticAI #AgentforceBuilder #Mvp-agentforce-prompt-builder-team
Nice find. From what I've seen, you shouldn't assume Agentforce tracks who changed a field between steps. Agents reason over what their actions return and what's in the conversation, so if nothing re-reads the record, the agent's view is whatever it last saw or did.
Patterns I'd use in Service Cloud:
1. Re-read before acting. Design actions (Flow or Apex) to query the current record state at the start of each step rather than relying on what the agent did earlier.
2. Make "who changed it" data. Turn on Field History Tracking for the key fields, and/or return LastModifiedById in your action, so the agent can see that a human edited the field. Then tell it in the instructions to treat a human edit as the source of truth.
3. Add a verification step. After the agent writes a value, have an action confirm the saved value matches what it intended before it moves on.
Your ActionTrace idea basically does #2 at the browser level. In the platform, the reliable version is to expose the change history through an action the agent can call.
