The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents a framework that ingests and processes untrusted data from evaluation sets (EvalSets) and conversation scenarios. This represents a potential indirect injection surface where malicious instructions in the test data could influence the agent being evaluated.
Ingestion points: EvalSet JSON files containing eval_cases, conversation, and conversation_scenario (SKILL.md, references/user-simulation.md).
Boundary markers: None specified in the documentation.
Capability inventory: The framework executes evaluation metrics, including LLM-as-judge and custom Python code (references/criteria-guide.md, references/multimodal-eval.md).
Sanitization: The documentation focuses on schema and implementation, without specific details on input sanitization.
[DYNAMIC_EXECUTION]: The skill describes a 'Custom Metrics' feature that enables the dynamic loading and execution of Python functions. Users can specify a module path in the eval_config.json file, which the framework then loads at runtime to perform evaluations (references/criteria-guide.md, references/multimodal-eval.md).
[EXTERNAL_DOWNLOADS]: The skill provides instructions to download and install the agents-cli utility using the uv package manager from official sources. It also references official documentation repositories hosted on Google's GitHub organization (SKILL.md, references/criteria-guide.md).

google-agents-cli-eval