[Release] google/adk-python v2.5.0: v2.5.0

Google ADK Python v2.5.0 introduces significant enhancements for agent development, including Cloud Run sandbox execution, MCP server integration, strict input validation, and improved workflow state management. Key additions include Voice Activity Detection support, HITL resumption capabilities, binary document parsing, and enhanced error handling callbacks. The release also addresses security concerns with module blocking, mTLS support, and comprehensive bug fixes across event compaction, model caching, and agent orchestration.

Key Points

• •Cloud Run sandbox option added to code executors for secure execution environments
• •MCP (Model Context Protocol) server integration: agents can now be served over MCP and support remote MCP servers with runtime callbacks
• •Strict input schema validation for LlmAgent workflow nodes ensures data integrity
• •Voice Activity Detection (VAD) events support in live mode for real-time audio processing
• •HITL (Human-in-the-Loop) resumption now works with standalone nodes and NodeTool for workflow interruption/resumption
• •State-based resumption for task-mode agent workflow nodes enables complex workflow recovery
• •Binary document parsing (DOCX, etc.) support via LoadArtifacts for expanded document handling
• •Non-blocking tools run in background tasks in live mode for improved responsiveness
• •Enhanced security: module blocking (posix, nt, _posixsubprocess, _socket) and mTLS support for API registry
• •Improved event compaction with language/tools preservation and rewound invocation exclusion

Workflow Diagram

Concepts