Instruction Adherence

Instruction adherence is an agent guardrail that evaluates how well your agents interpret subagent instructions when generating responses.

Instruction Adherence Score

Instruction adherence evaluates the agent's response and scores it based on how well the agent interprets and follows the subagent instructions. There are three score categories.

Score	Description
High	Indicates that the agent interprets and fully follows the subagent instructions, addressing key points and providing the required information.
Low	Indicates that the agent doesn’t interpret or follow the subagent instructions accurately. It fails to follow at least one instruction, leading to an incorrect response.
Uncertain	Indicates that instruction adherence can't be conclusively determined. This can be due to an ambiguous response or action, incomplete response, or conflicting interpretations of the subagent instructions.

Scores and Explanations Are Logged in Data 360 and Visible in Agent Event Logs

Instruction Adherence scores are logged in the Einstein Generative AI Audit and Feedback data model in Data 360, which must be turned on in your org (see Set Up Einstein Generative AI Audit and Feedback). Scores are visible in agent event logs. Turn on Enhanced Agent Event Logs.

In addition to the scores, an explanation of why an agent's response was classified as high, low, or uncertain is included. You can use this information to see how your agent is following the subagent instructions. It can also help you determine whether the subagent instructions need improvement or more instructions. See Best Practices for Writing Subagent Instructions and Troubleshooting Agents.

Note This feature is supported only in English and is available only for Agentforce for Service agents.