Apply Guardrails at Runtime
Observing, measuring, and testing are essential, but when an answer could harm a patient, you need to act in the moment. Guardrails (agent controls) evaluate each step at runtime and either block it or steer the agent toward a safe response.
Persona
As Careful Health Provider’s AI engineer, you’ve now seen the assistant produce ungrounded
medical advice and call a sensitive deletion tool. Detection isn’t enough; you need to stop
the dangerous deletion and redirect the unsafe answer, in production, without taking the
agent offline.
Agent Control is your single watchtower across every agent. Enforce control
policies with runtime blocking and steering, and push policy changes that stop unwanted behavior in seconds, without redeploying or taking agents offline.
Block vs. steer
- Block: the step is rejected outright. In the app this surfaces as a
ControlViolationError, and the user receives a safe “this action was blocked” message. Use it for the irreversible patient-record deletion. - Steer: the step is sent back for revision with guidance. In the app this surfaces as a
ControlSteerError; the agent retries using the steering guidance before giving up. Use it to redirect an unsafe medical answer into a safe one rather than refusing the user entirely.
Where to work
This chapter works in
~/workshop/healthcare-assistant/4-app-with-controls. It builds on the
instrumented app and adds the Agent Control SDK, configuration, and the @control-decorated
steps. The folder ships complete, so it doubles as the reference for this chapter.Continue to the subsections to add Agent Control, define controls, and test block and steer.