You are building a custom customer service agent and have compiled a golden dataset of historical user questions and their ideal answers. Before deploying the agent, you need to programmatically evaluate both the relevance of retrieved documents and the groundedness of the agent's generated responses. You want to minimize custom evaluation code.
What should you do?

• A. Deploy the agent to Agent Runtime, run the agent against the golden dataset, and review the metrics using Google Cloud Observability.
• B. Execute the evaluation of the golden dataset against the agent using ADK eval set tooling, and configure LLM-based autorater metrics.
• C. Run a batch evaluation job against the agent using the golden dataset to calculate BLEU and ROUGE scores for the generated responses.
• D. Configure a continuous evaluation pipeline using the Agent Search ranking API to compare the semantic similarity between the agent's generated response and the golden dataset's ideal answer.

You manage an agent built with Agent Development Kit (ADK) that has six tools. You discover that the agent's 95th percentile response time has doubled over the last 7 days. Dashboards show the increased time but not its source. One of the six tools calls an internal pricing service maintained by another team in your organization. The service reports its own internal timings to the same observability backend, preserving the caller's request context.
You need to identify the step responsible for the added time, including any work performed inside the pricing service.
What should you do?

• A. Correlate the agent and pricing service logs in Cloud Logging by trace identifier.
• B. Read the agent's OpenTelemetry traces in Cloud Trace as a span waterfall.
• C. Compare the gen_ai.tool.execution.duration metric for the tools in Cloud Monitoring.
• D. Enable Cloud Profiler for the agent service.

You are implementing a custom skill in the Antigravity CLI to automate service deployments. When tasks run in Agent Mode, the CLI does not detect the skill and skips user confirmation before executing deployment commands. You want to enable effective global skill discovery while ensuring human approvals are required for high-risk actions.
What should you do?

• A. Verify that a SKILL.md file containing valid YAML metadata is present in the ~/.gemini/antigravity-cli
  /skills/directory, and execute the Agents CLI without auto-approval flags.
• B. Upload the deployment script to Cloud Deploy as an agent extension, and run the agents-cli setup command with the -- force flag.
• C. Register the skill as an MCP server inside the .agents/skills/directory for each deployment, and launch the Agents CLI using the -y flag.
• D. Verify that a SKILL.md file containing valid YAML metadata is present in the .agents/skills directory, and execute the Agents CLI without auto-approval flags.

You are implementing a continuous evaluation pipeline for an agent built with Agent Development Kit (ADK) that invokes multiple external API tools. During evaluation runs, the agent often selects the wrong tools or executes them in the wrong order, causing errors in multistep workflows. You need to establish automated evaluation gates in your CI/CD pipeline to improve execution reliability before promoting model or prompt changes to production.
What should you do?

• A. Configure Agent Gateway policies with Model Armor templates to inspect function-calling payloads at runtime trajectories and automatically trigger fallback routes when unexpected tool call arguments are detected.
• B. Configure continuous evaluation runs in the CI/CD pipeline using the gen AI evaluation service configured with trajectory metrics to evaluate the agent's tool call sequence against golden datasets.
• C. Configure Agent Platform Memory Bank session persistence and Agent Registry governance rules to replay historical user session trajectory and automatically downgrade agent revisions when tool call errors occur.
• D. Configure automated presubmit tests in Cloud Build using the gen AI evaluation service configured with response-level metrics to score the final text generated by the agent.