Agent Guardrail Escalation App

Escalation template purpose built for Agent Guardrail escalation actions

This template, once published and deployed, can be added as an escalation action in the Agent Guardrail builder. You can access this Guardrail builder by adding any guardrail (custom, PII detection, Prompt injection detection), then selecting the 'escalate' action at the bottom and referencing the deployed app.

Depending on the scope of the guardrail (agent, LLM call, or tool), this offers the ability to have agent inputs/outputs, LLM request, tool inputs/outputs reviewed by an assignee in Action Center before the agent/LLM/tool call happens AND/OR before the agent/LLM/tool output is sent back to the agent/user. For sensitive business applications, workflows and agents, users can identify specific circumstances under which a human in the loop review needs to take place. The escalation will happen deterministically during the agent execution.

Important notes

• The template can be further customized, but cannot have existing schema removed (additions only).
• The app must be published and deployed in Orchestrator with an unattended license and serverless machine assigned to the folder. Please consult the Orchestrator documentation for further support.
• Modifications to the inputs and outputs back to the agent can result in errors. Please modify with care so as not to cause any undesired agent errors or tool calls.

• Review agent inputs before execution starts
• Review LLM requests before the LLM call is made
• Review tool inputs before a tool call is made
• Review tool outputs before sending back to the agent context
• Review agent result before sending back to the user
• Make modifications to inputs or outputs if desired
• Add a reason for accepting/denying/modifying the call for logging purposes