Workflow routing and real-time voice
Google released version 2.13.0 of the Agent Development Kit, bringing dedicated routing models and spoken dialogue capabilities to agent graphs. Workflows can now route tasks through TypesafeClassifier and ClassifierNode, which rely on TypeSafe’s Jev models. Google states these models calculate calibrated probabilities in a single forward pass to make routing decisions faster than a standard model query.
The framework also introduces live voice interactions through CascadeLiveConnection. The system integrates Gemini speech-to-text and text-to-speech pipelines while supporting conversational barge-in, allowing a user to interrupt an ongoing audio response without breaking session state.
Session storage, memory, and telemetry shifts
For distributed agent deployments, the update introduces experimental MongoDB integrations. The new MongoDbSessionService allows teams to persist execution states and share active sessions across multiple runner replicas. In addition, MongoDbMemoryService adds long-term recall using keyword search over past turns across standard database clusters, removing the operational need for vector embeddings or dedicated vector indexes.
Alongside stabilizing twelve earlier experimental features, the release applies breaking changes to telemetry pipelines. ADK now defaults to the OpenTelemetry GenAI v2 schema, discarding the call_llm span unless operators manually set an environment variable to retain the legacy v1 format.
Our assessment
Teams building interactive voice agents should evaluate whether the barge-in implementation properly halts queued tool executions when interrupted by a user. For workflow orchestration, replacing prompt-based LLM routers with Jev classifier nodes may lower latency, but developers must verify classification accuracy on complex intents. Teams maintaining observability pipelines will need to update ingestion collectors to match the new OpenTelemetry GenAI schema before upgrading.
Source: Google original publication.
Read the original material
Google (opens in a new tab)The article date belongs to Agentic Horizon. The source publication date is listed separately. Vendor claims remain attributed; we have not independently tested the reported capability.
