The Agent Skills Directory

[COMMAND_EXECUTION]: The skill implements desktop and shell automation capabilities.
The ComputerUseAgent class uses pyautogui for mouse clicks, typing, and key presses in the perception-reasoning-action loop in SKILL.md.
The AnthropicComputerUse implementation includes a bash tool using subprocess.run to execute system commands such as scrot for screen capture in SKILL.md.
[INDIRECT_PROMPT_INJECTION]: The skill defines vision-based agents that process screen content, creating a surface for instructions embedded in visual data to influence agent behavior.
Ingestion points: Screen state is captured via pyautogui.screenshot() and the scrot command line tool in SKILL.md.
Boundary markers: The provided implementation snippets do not include explicit input validation or boundary delimiters for processed visual data.
Capability inventory: The agent has broad system access via pyautogui (UI control) and a bash tool (shell access) in SKILL.md.
Sanitization: No specific sanitization or filtering is applied to the vision-based inputs or command outputs in the example code.
[EXTERNAL_DOWNLOADS]: The provided Dockerfile includes patterns for installing software from official package registries.
The configuration installs system packages via apt-get and Python packages via pip3 install from the requirements file in SKILL.md.

computer-use-agents