The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes a shell script (create-agent.sh) used to scaffold new agent projects. Furthermore, the agents created by the SDK are frequently configured with the Bash tool, enabling them to execute arbitrary shell commands to perform tasks such as data analysis or system interaction.
[REMOTE_CODE_EXECUTION]: The skill explicitly documents an 'Advanced Pattern' called 'Code Execution with MCP' in SKILL.md. This pattern instructs the AI agent to write its own Python scripts and execute them locally via the Bash tool to process large datasets or perform complex logic. This represents a high-privilege capability where code is generated and executed at runtime.
[EXTERNAL_DOWNLOADS]: The documentation and scaffolding scripts guide the user to download and install the @anthropic-ai/claude-agent-sdk and other development dependencies (tsx, typescript) from official npm registries. These references target a trusted organization and are considered standard development practice.
[PROMPT_INJECTION]: The skill facilitates the creation of agents that ingest untrusted user input and external data (via Read or Bash tools) and possess powerful capabilities (Write, Edit, Bash). This creates a significant surface for indirect prompt injection.
Ingestion points: The query() function in agent.ts accepts raw user prompts and processes tool outputs from the filesystem or shell.
Boundary markers: The default agent templates and system prompts do not include explicit delimiter markers (like XML tags) to separate instructions from data.
Capability inventory: Agents created via the scaffold are granted access to Read, Grep, and Glob by default, with instructions encouraging the addition of Write, Edit, and Bash for automation.
Sanitization: There is no built-in sanitization or validation of external content before it is interpolated into the agent's context in the provided templates.

agent-sdk