The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user and agent to install the 'a-evolve' package via pip and clone its source code from 'github.com/A-EVO-Lab/a-evolve'. This organization and repository are not recognized as trusted vendors or well-known services.
[REMOTE_CODE_EXECUTION]: The framework uses an LLM-driven 'Evolution Engine' that performs workspace mutations. This process involves a remote model generating instructions and skill procedures that are subsequently loaded and executed by the local agent environment.
[COMMAND_EXECUTION]: The default evolution engine ('SkillForge') is documented to use bash tool access to directly edit files within the agent workspace (including prompts and scripts). This capability allows the LLM to modify local file state and execute shell commands to manage the evolution loop.
[PROMPT_INJECTION]: The skill features an inherent vulnerability to indirect prompt injection. The evolution process ingests untrusted data from benchmarks and task feedback to mutate the agent's system prompt and skills. * Ingestion points: Benchmark task inputs ('benchmark.get_tasks') and evaluation feedback ('benchmark.evaluate') are processed as observations. * Boundary markers: No explicit boundary markers or delimiters are described to separate untrusted task data from instructions during the prompt mutation phase. * Capability inventory: The 'AgentWorkspace' provides APIs for writing prompts and skills, and the engine has access to bash tools. * Sanitization: No sanitization or validation of the ingested observation data is documented before it is passed to the LLM for analysis and mutation.

evolving-ai-agents