The Agent Skills Directory

[PROMPT_INJECTION]: The skill defines patterns for evaluating and tracing agent responses which process untrusted user input (Indirect Prompt Injection). This represents a vulnerability surface inherent to agent monitoring, but the skill specifically mitigates this by instructing users to implement MLflow's built-in safety scorers and guidelines adherence metrics.
Ingestion points: Agent request objects in ResponsesAgent.predict() and evaluation_data processed by mlflow.genai.evaluate().
Boundary markers: Encourages the use of GuidelinesAdherence with explicit behavioral constraints for LLM-based evaluation judges.
Capability inventory: The skill utilizes MLflow's standard model logging and tracing capabilities, and the Databricks SDK for metadata management.
Sanitization: Promotes the use of the ResponsesAgent framework which enforces structured schema validation for inputs and outputs.
[EXTERNAL_DOWNLOADS]: The skill references established, well-known libraries including mlflow, pandas, and databricks-sdk in its configuration snippets and requirements. These originate from official, trusted sources within the Databricks ecosystem and do not constitute an external dependency risk.

mlflow-genai-foundation