QuestionQ55

VerifiedSecuring and governing agentic workflows

You are designing a human-in-the-loop architecture for an enterprise agent built with Agent Development Kit (ADK) that executes high-risk business operations. You must pause agent execution during sensitive tool invocations, present approval requests out-of-band without risk of prompt injection bypass, and reliably resume execution upon human confirmation.

What should you do?

Explanation

Policy-based interception can pause sensitive tool calls for human approval. An authenticated approval outside the model’s chat stream, matched to the confirmation call ID, avoids relying on the model to interpret an approval. Google documents this pattern using SecurityPlugin and a policy engine. (github.com)

The intended choice has an important limitation: ADK’s action-confirmation documentation lists VertexAiSessionService as unsupported. A production implementation therefore requires a supported persistence and approval design rather than relying on that combination as written. (google.github.io)

Learn more

Community Discussion

No comments yet. Be the first to start the discussion!