The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents interfaces for processing untrusted data, creating potential injection surfaces.
Ingestion points: Data enters the agent context through email handlers (onEmail in references/email.md), webhook endpoints (onRequest in references/webhooks-push.md), and web scraping tools (browser_execute in references/browse-the-web.md).
Boundary markers: While using standard SDK message conversion utilities, the documentation does not demonstrate explicit delimiting of external content within prompts.
Capability inventory: Agents built using these patterns have access to SQL databases (this.sql), sandboxed JavaScript execution (codemode), and browser control.
Sanitization: The documentation explicitly covers security best practices such as signature verification for webhooks and secure HMAC-based reply routing for emails to ensure data authenticity.
[DYNAMIC_EXECUTION]: The documentation includes experimental features designed for runtime code generation and execution by the agent.
Evidence: The Codemode feature (references/codemode.md) enables LLMs to write JavaScript that executes in an isolated Worker sandbox using DynamicWorkerExecutor.
Evidence: The browser_execute tool (references/browse-the-web.md) allows LLMs to script browser interactions using CDP.
[EXTERNAL_DOWNLOADS]: The skill refers to several standard and official packages for installation via npm.
Evidence: Instructions include installing agents, @cloudflare/codemode, and ai from the public registry to support SDK functionality.

agents-sdk