Google ADK Python v2.13.0
Google ADK Python v2.13.0 adds TypeSafe Jev model classifiers, live audio barge-in with Gemini STT/TTS, experimental MongoDB session and memory services, and breaking OpenTelemetry schema updates.
Why this signal matters
Google has released ADK Python version 2.13.0 with significant new integrations, performance enhancements, and breaking telemetry changes. Key additions and modifications include: - Breaking Telemetry Changes: Defaults now follow OpenTelemetry GenAI v2 ( otel semconv 1 44 ) semantics, deprecating and removing the call llm span. Developers requiring the legacy schema can opt in using ADK TELEMETRY SCHEMA VERSION OPT IN=otel semconv 1 36 . - Workflow Classification: Introduced TypesafeClassifier (for TypeSafe Jev models) and ClassifierNode , designed for routing workflows with calibrated probabilities faster than standard LLMs. Added general abstractions including BaseClassifier and LlmClassifier . - Live Audio Support: Added natural real-time voice interactions via CascadeLiveConnection , featuring barge-in handling and integration with Gemini speech-to-text (STT) and text-to-speech (TTS). - MongoDB Integrations: Added experimental MongoDbSessionService (enabling cross-runner session sharing) and MongoDbMemoryService (keyword-based turn recall without embedding models or vector indexes). Added capability to deploy MongoDbToolset to Agent Engine without connection strings, leveraging Atlas Automated Embedding. - Performance & Stability: Execution turns for string-model agents are noted to be up to 8x faster with decreased event-loading overhead. Twelve experimental features have transitioned to stable status, and an OpenSandbox environment has been added.
Actionable summary
ADK Python v2.13.0 introduces a breaking change by switching the default telemetry schema to OpenTelemetry GenAI v2 ( otel semconv 1 44 ), dropping the call llm span. Key functional updates include fast workflow routing via TypesafeClassifier and ClassifierNode , live real-time audio interaction with barge-in support and Gemini STT/TTS, and MongoDB-backed session and non-vector memory integrations.
- Agent usefulness
- 89/100
- Confidence
- 96%
- Canonical data
- JSON + Markdown
What builders should check
- Update telemetry parsing and downstream observability systems to support OpenTelemetry GenAI v2 semconv 1.44, or set ADK TELEMETRY SCHEMA VERSION OPT IN=otel semconv 1 36 to retain the legacy v1 schema.
- Remove references to the deprecated call llm span in telemetry dashboards or processing pipelines.
- Review workflow routers to consider switching from LLM-based prompting to TypesafeClassifier or ClassifierNode for reduced latency.