The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The SDK processes user-generated messages and tool outputs, which are standard ingestion points for data that could contain instructions.
Ingestion points: Data enters the system via agent.sendMessage (SKILL.md) and agent.submitToolResult (references/tool-builder.md).
Boundary markers: The provided code examples do not explicitly demonstrate the use of delimiters to isolate untrusted input.
Capability inventory: The SDK supports powerful tools such as uploadFile for file system interaction and webhookTool for network requests.
Sanitization: Sanitization logic is omitted from basic documentation examples, as it is expected to be managed by the application logic or the platform's core model guardrails.
[DYNAMIC_EXECUTION]: The documentation includes features and examples for executing code at runtime to enable agent autonomy.
The references/tool-builder.md file demonstrates a tool handler using eval() to process mathematical expressions.
The references/agent-patterns.md file describes an internal tool capability .codeExecution(true) which allows agents to generate and run code to solve problems.

javascript-sdk