The Agent Skills Directory

[COMMAND_EXECUTION]: The skill documents the implementation of shell and command-line capabilities via agent.enable_shell. It provides a detailed framework for security controls, including filesystem root restrictions, environment isolation using Docker or Windows Sandbox, and mandatory user approval policies for command execution.
[DYNAMIC_EXECUTION]: The skill defines protocols for executing model-generated code in Python, Node.js, and other languages. It explicitly mandates the use of isolated execution providers such as gVisor or Docker and includes instructions to 'fail closed' if required isolation levels cannot be met.
[EXTERNAL_DOWNLOADS]: The skill contains instructions for installing the agently and agently-devtools packages from official registries and describes the process of pulling Docker images for code execution. These downloads are directed toward trusted vendor-owned resources.
[PROMPT_INJECTION]: Analysis of the skill's instructions regarding console display and logging filters confirms they are benign UI optimizations designed to improve readability by suppressing redundant heartbeats and progress events, while maintaining a full audit trail in the backend DevTools system.

agently-runtime