The Agent Skills Directory

Dynamic Code Execution: The skill documents the use of custom_function and custom_function_file to perform evaluations using Python scripts. These functions execute in the local environment where the CLI is run. The documentation includes a clear notice that these scripts run with the user's local privileges and recommends using only trusted code for this purpose.\n- Indirect Prompt Injection Consideration: As an evaluation tool, the skill processes agent trajectories and user-provided datasets which serve as input for LLM-based judges and custom metrics. This represents a potential surface for indirect prompt injection if the analyzed content contains malicious instructions. The documentation mitigates this by providing structured schemas and rubric-based evaluation methods.\n- External Tool Installation: The guide facilitates the installation of the official google-agents-cli package. This is a standard procedure for utilizing the vendor's specialized evaluation tools and does not involve untrusted third-party sources.\n- Network and File Operations: The skill describes workflows that involve sending data to well-known Gemini platform endpoints for analysis and writing evaluation results to the local filesystem. These operations are core to the tool's intended purpose and are transparently documented.

google-agents-cli-eval