Agent guardrail monitoring
Key Takeaway
Monitor two agent guardrails. Instruction adherence scores each agent response as High, Low, or Uncertain, based on whether the agent followed its subagent instructions. Task resolution scores each conversation as Fully Resolved, Partially Resolved, or Unresolved, based on whether the user's task got done.
Agent guardrail monitoring is how you check whether Agentforce agents behave well in real conversations. Two agent guardrails score agent behavior. Each one sorts agent behavior into score categories and logs the scores and explanations in Data 360. You review those results, report on them, and use them to improve the agent's instructions.
Two guardrails, two questions
The two guardrails answer different questions. Instruction adherence scores the agent's response. Task resolution scores the whole conversation.
If you want to know whether the agent followed its subagent instructions when it generated a response...
Monitor instruction adherence.
If you want to know whether a conversation resolved what the user asked for...
Monitor task resolution.