Apply Guardrails at Runtime

Add Agent Control to the App

8 minutes

First, wire the Agent Control SDK into the app: add the configuration, install the packages, and register the controllable steps.

Exercise Add Agent Control
1

Set up the environment

Change into the agent controls folder:

bash 
cd ~/workshop/healthcare-assistant/4-app-with-controls
2

Create an Agent Control Config Map

Run the following command to create a Kubernetes config map, which the application will use to configure Agent Controls:

bash 
kubectl create configmap galileo-agent-control-config \
  --from-literal=GALILEO_API_URL="https://api.multitenant.galileocloud.io" \
  --from-literal=AGENT_CONTROL_URL="https://console.multitenant.galileocloud.io/api/agent-control" \
  --from-literal=AGENT_CONTROL_AGENT_NAME="agent-control-example" \
  --from-literal=AGENT_CONTROL_API_KEY_HEADER="Galileo-API-Key" \
  --from-literal=AGENT_CONTROL_RUNTIME_AUTH_MODE="jwt" \
  --from-literal=AGENT_CONTROL_TARGET_TYPE="log_stream"
3

Add the Agent Control packages

Confirm requirements.txt includes the Agent Control SDK and evaluators:

text 
agent-control-sdk[galileo]>=7.10.0
agent-control-evaluators>=7.10.0
agent-control-evaluator-galileo>=7.10.0

Note about the SDK

This workshop was built using the older agent-control packages. For new deployments, we recommend using the following packages instead:

text 
agent-control-sdk[splunk-ao]>=7.10.0
agent-control-evaluators>=7.10.0
agent-control-evaluator-splunk-ao>=7.10.0

Refer to the Agent Control document for details about this newer SDK.

4

Add the imports and decorate the steps

In agent.py, we’ve added the Agent Control imports at the end of the import section:

python 
from agent_control import ControlSteerError, ControlViolationError, control

The controls stage uses these in three places (already wired up in this folder):

  • The LLM call is wrapped with @control(step_name=LLM_STEP_NAME), where LLM_STEP_NAME = "Healthcare Assistant", so the model’s responses can be evaluated, blocked, or steered.
  • Each tool is registered as a controllable step (get_patient_info, delete_patient_record, and a shared retrieval_step for search tools) via the helpers in helpers/agent_control_helpers.py.
  • The agent enables control spans on the Galileo logger (galileo_logger.enable_agent_control()) and registers the steps with init_agent_control(...) so the console knows which steps exist for this agent.

How block and steer are handled in code

When a control fires, the SDK raises an exception that the agent catches:

  • ControlViolationError → the step is blocked; the user gets a friendly blocked message.
  • ControlSteerError → the step is steered; the agent rebuilds the prompt with the steering guidance and retries (up to MAX_STEER_RETRIES) before falling back to a safe message.

What’s the difference between how the app handles a ControlViolationError and a ControlSteerError?

Click here to see the answer
A ControlViolationError blocks the step: the action is stopped and the user receives a friendly blocked message. A ControlSteerError steers the step: the agent retries with the steering guidance appended to the prompt (up to a few attempts) to produce a corrected response, only falling back to a safe message if it still can’t comply.