The Agent Skills Directory

[Command Execution]: The skill provides instructions for running agents that may execute shell commands. It proactively recommends the use of safety flags and sandbox microVMs to ensure commands are either reviewed or isolated from the host system.
[External Downloads]: Documentation regarding prerequisites like the sbx CLI and docker-agent plugin points to official vendor repositories and standard package managers. These resources originate from the vendor's own infrastructure and are standard for the tool's operation.
[Indirect Prompt Injection Surface]: The skill manages environments where agents process potentially untrusted input through configuration files and task prompts.
Ingestion points: Agent configuration files (e.g., agent.yaml) and user-supplied tasks passed via the --exec flag.
Boundary markers: The skill describes runtime safety modes (strict, balanced, restricted) which act as policy-based boundaries for tool approval.
Capability inventory: Shell command execution and filesystem access via the agent's toolsets are the primary capabilities identified.
Sanitization: Sandbox isolation (--sandbox) is detailed as a mechanism to isolate the agent's filesystem and process activity within a microVM, preventing direct impact on the host system.

docker-agent-run