The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill's documentation and production examples demonstrate the use of exec() to evaluate dynamically generated Python agent architectures. Specifically, the evaluate_agent function takes a candidate string (generated code) and executes it directly using exec(candidate, exec_globals). While presented as an example, this pattern is inherently risky as it allows arbitrary code execution within the agent's environment.
[COMMAND_EXECUTION]: The evaluate_solver production example demonstrates executing generated Python code using subprocess.run(['python', '-c', candidate], ...). This allows the execution of arbitrary shell commands if the generated candidate contains malicious logic.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external datasets (dataset, valset) and natural language inputs (objective, background) to guide optimization. These ingestion points lack explicit boundary markers or sanitization instructions in the provided documentation, creating a surface where instructions embedded in the data could influence the agent's behavior during the reflective evolutionary search process.
[EXTERNAL_DOWNLOADS]: The skill requires the installation of the gepa package from an external source (PyPI) and references documentation/source code on GitHub. These are documented neutrally as required dependencies for the skill's functionality.

dspy-optimize-anything