Google ADK (Python) v2.13.0

Summary

ADK for Python 2.13.0 is dated 2026-10-22 in its changelog (the GitHub release was created on 2026-10-23). It promotes 12 experimental features to stable and adds classifier nodes for routing workflows. It also adds barge-in for live audio, MongoDB session and memory services, and up to 8× faster turns for agents given a model name as a string. It has one breaking change in telemetry.

Key points

• Breaking: telemetry now defaults to the OpenTelemetry GenAI v2 conventions (otel_semconv_1_44), and the call_llm span is gone. To keep the old schema, set ADK_TELEMETRY_SCHEMA_VERSION_OPT_IN=otel_semconv_1_36.
• Classifiers for routing: BaseClassifier, LlmClassifier, TypesafeClassifier (uses TypeSafe's Jev models, which return calibrated probabilities faster than an LLM), and ClassifierNode for workflow graphs.
• Memory and sessions: delete_memory on BaseMemoryService, plus experimental MongoDbSessionService and MongoDbMemoryService. The memory service uses keyword matching and needs no vector index.
• Live audio: barge-in for CascadeLiveConnection, and Gemini speech-to-text and text-to-speech.
• Evals and telemetry: invocation time is split into model, tool, and framework time. Eval results include Vertex rubric verdicts. AgentEvaluator and the CLI eval command exit with a pass/fail code, which is useful in CI.
• Sandboxes and deployment: an OpenSandbox environment, AgentEngineEnvironment, and readiness helpers in the sandbox manager.
• MCP: only fixes (auth and header precedence, McpToolset now uses McpTool). The notes do not say that MCP SDK 2.x is now the default, so it still seems to be opt-in.
• Stable features: the release notes do not list the 12 promoted features by name. They point to the feature registry guide.

Why it matters

Exit codes that match eval results make ADK evals easy to use as a CI gate. Teams with dashboards built on the old call_llm span must change them or set the opt-in variable before upgrading.