You are configuring an agentic workflow in Agent Studio for an enterprise customer analytics platform. The agent was initialized with default system instructions, prompt templates, and a Gemini Flash model family. In staging, you notice that the agent frequently hallucinates when synthesizing data across multiple knowledge documents and produces inconsistent output formats.
You need to ensure that the agent reliably uses verified source context and follows a standardized response format.
What should you do first?

• A. Update the system instructions to route date parsing requests to a custom Cloud Run OpenAPI extension tool, and reconfigure prompt templates to increase chunk overlap in Agent Search.
• B. Update the system instructions to restrict answers strictly to retrieved source data, and reconfigure prompts with chain-of-thought reasoning steps and few-shot input-output examples.
• C. Update the system instructions to reduce the temperature parameter and maximum top-p sampling to zero, and change the model to a Gemini Pro model family.
• D. Update the system instructions with explicit data redaction rules, and configure prompt templates to send all output through a Model Armor filter for schema enforcement.

You are configuring a secure agent-to-tool API integration for a financial agent on Gemini Enterprise Agent Platform. The agent calls an internal payment processing microservice protected by OAuth 2.0.
To comply with strict enterprise security policy, you must ensure that:
OAuth client credentials used for service-to-service tool authentication are encrypted at rest and never hardcoded in tool definitions or repository source code.
Tool execution is strictly limited to the minimum required API permissions, so the agent cannot perform unauthorized administrative operations on the payment microservice.
What should you do?
Choose two

• A. Configure the tool definition to use OAUTH2_IMPLICIT_GRANT with wildcard scopes (*).
• B. Define explicit OAuth scopes such as payments.read and payments.create in the auth provider definition, and bind the provider to the specific payment tool definition in Agent Runtime.
• C. Grant the roles/owner IAM role on the payment microservice project to the agent's identity.
• D. Encode the OAuth client secret in Base64, and pass the secret in the tools.py environment variable map.
• E. Store the OAuth client credentials in Secret Manager, and reference the Secret Manager resource URI within the Agent Identity auth manager provider configuration.

You are designing a multiagent enterprise architecture on Google Cloud using the Agent2Agent (A2A) protocol. A central supervisor agent delegates complex subtasks to independently deployed, specialized subagent microservices hosted in separate projects. During high-volume batch runs, long-running subagent tasks fail because of connection timeouts caused by the supervisor agent blocking synchronous HTTP polling.
You need to rearchitect the interagent communication layer.
What should you do?

• A. Configure a Pub/Sub topic between projects and instruct the supervisor to issue system-level SIGCONT process signals to remote subagent container runtimes upon task completion.
• B. Increase the supervisor agent's HTTP client connection timeout to 3,600 seconds, and set a stream=True parameter in the request payload body.
• C. Configure the subagents to expose A2A-compliant server-sent event (SSE) streaming endpoints and event webhooks.
• D. Wrap the subagent microservices in an MCP HTTP transport layer using Cloud Run container instances.

You are developing a custom financial analysis agent in Python with Agent Development Kit (ADK). The agent calls an exchange rate function when users request cross-border portfolio valuations, but in multi-turn interactions, the model often passes invalid currency codes or fails to trigger the function. You need to ensure that the agent reliably invokes the function with valid parameter types.
What should you do?

• A. Define tool parameter schemas as unformatted string lists inside the agent system instructions.
• B. Define explicit Python type annotations and detailed parameter docstrings in the tool function signature.
• C. Apply a Pydantic @validate_call decorator parameter to the Python function to raise a runtime ValidationError result if an unapproved parameter value is passed to the function.
• D. Implement a custom retry-loop wrapper around the tool execution call to catch malformed parameter exceptions and automatically reprompt the model.

You are deploying an enterprise conversational agent in Google Cloud that connects to sensitive human resources (HR) databases. Your organization's compliance policy requires all user queries and model completions to be filtered in real time to prevent prompt injection attacks, block harmful content, and sanitize personally identifiable information (PII) before responses reach the user interface. You need to implement a managed security enforcement layer that meets these requirements with minimal latency.
What should you do?

• A. Configure Model Armor filter templates with linked Sensitive Data Protection de-identification templates, and attach them to each agent's runtime.
• B. Enforce Model Armor filter templates with linked Sensitive Data Protection de-identification templates at the Agent Gateway for all agent traffic.
• C. Apply a Sensitive Data Protection de-identification template at the Agent Gateway, and filter for injection and harmful content using Model Armor's input and output templates
• D. Deploy a custom Llama Guard model on Model Garden behind the Agent Gateway, and add a Sensitive Data Protection inspection template for PII.