The Agent Skills Directory

[Documentation of Execution Capabilities]: The skill provides instructions for building agents capable of executing Python functions, tools, and dynamic graph nodes. While these features enable powerful agent behavior, they are presented within the context of a legitimate development framework and include guidance on error handling and validation.
[Indirect Prompt Injection Surface]: The documentation describes how to build agents that process external data, such as tool outputs and user messages. This represents a potential attack surface where untrusted data could influence agent behavior.
Ingestion points: Agents ingest data via new_message in runners, node_input in workflow nodes, and results from external tools (MCP, OpenAPI).
Boundary markers: The skill recommends using instruction templates with {var} placeholders and structured data formats.
Capability inventory: Built agents may have access to the file system (via MCP filesystem server), network resources (via OpenAPI toolsets), and data warehouses (via BigQuery toolsets).
Sanitization: The skill strongly advocates for using Pydantic BaseModel for all node inputs, outputs, and state values to ensure schema-level validation and coercion.
[Standard External Dependencies]: The skill references the installation of the google-adk library and associated development tools (e.g., pytest, mcp). These references target the vendor's official packages and established development registries.
[Credential Management Best Practices]: The skill provides clear instructions on managing sensitive information, such as Google API keys, by using .env files and specifically advising against committing these files to version control.

adk-agent-builder