Google ADK for Python v2.13.0 adds MongoDB sessions and memory, live audio barge-in

Google has released version 2.13.0 of ADK for Python, moving 12 experimental features to stable status. The update adds TypesafeClassifier and ClassifierNode, which let workflows route requests using TypeSafe's Jev models. According to the GitHub release, Jev returns calibrated probabilities in a single forward pass and is much faster than calling an LLM.

MongoDbSessionService persists sessions in MongoDB and shares them across runner replicas. MongoDbMemoryService adds long-term memory that recalls past turns by keyword match, so no embedding model or vector index is needed. Both are experimental, and MongoDbToolset can now be deployed to Agent Engine without packaging a connection string.

Live sessions gain barge-in support in CascadeLiveConnection, along with Gemini speech-to-text and text-to-speech integration. Telemetry now defaults to OpenTelemetry GenAI v2 and drops the call_llm span, while setting ADK_TELEMETRY_SCHEMA_VERSION_OPT_IN=otel_semconv_1_36 keeps the legacy v1 schema.

Teams whose dashboards or alerts depend on the call_llm span or the v1 GenAI schema should set the opt-in variable or update their queries before upgrading. Agents that route with an LLM can evaluate TypesafeClassifier as a faster alternative for workflow routing. MongoDB-backed sessions let several runner replicas serve the same agent state, though the MongoDB services are still experimental.