adk-python 2.13.0 promotes 12 features to stable, adds MongoDB

Google published adk-python v2.13.0 on October 23, 2026 (the changelog heading is dated October 22). It is a stabilization release: 12 experimental features move to stable, and the release adds MongoDB-backed sessions and memory, classifier nodes for workflow routing and barge-in for live audio. It also changes the default telemetry schema, which is the one item to check before upgrading.

Release facts

What changed

• 12 features promoted to stable. The release notes say 12 mature experimental features are now stable, and a new ADK_SUPPRESS_EXPERIMENTAL_FEATURE_WARNINGS setting is honored for the ones that remain experimental.
• MongoDB sessions and memory. MongoDbSessionService persists sessions in MongoDB so several runner replicas can share them. MongoDbMemoryService adds long-term memory that recalls past turns by keyword match, with no embedding model or vector index. Both are marked experimental.
• MongoDB on Agent Engine. MongoDbToolset can be deployed to Agent Engine without packaging the connection string, and Atlas Automated Embedding can embed search queries server-side.
• Classifiers for routing. New BaseClassifier, LlmClassifier and ClassifierNode let a workflow branch on a classification step. A TypesafeClassifier integration uses TypeSafe's Jev models, which the notes describe as returning calibrated probabilities in a single forward pass.
• Live audio. CascadeLiveConnection gets barge-in support, plus Gemini speech-to-text and text-to-speech.
• Performance. The notes claim up to 8x faster execution turns for agents configured with a model string, and less event-loading overhead.
• Evaluation and tooling. Eval results split invocation time into model, tool and framework time, adk eval now exits non-zero on failures, and Vertex rubric verdicts show up in eval results.
• Security fix. RestApiTool gets SSRF protection.

The breaking change, in practice

If you export ADK traces to a dashboard that keys on the call_llm span or the 1.36 GenAI attributes, that view will change after the upgrade. The release notes give the escape hatch: set ADK_TELEMETRY_SCHEMA_VERSION_OPT_IN=otel_semconv_1_36 to keep the legacy schema while you migrate queries. The release also caps litellm below 1.106.3, which matters if you pin LiteLLM yourself.

Who should upgrade now

Teams running ADK across several replicas get the most from this release, because shared MongoDB sessions remove a common reason to write a custom session store. Teams with OpenTelemetry dashboards should upgrade on a branch first. If you only use ADK locally with in-memory sessions, the stable-feature promotions and the adk eval exit code fix are the practical wins.

Questions

Does adk-python 2.13.0 need a vector database for long-term memory? Not for the new MongoDB memory service. The release notes say MongoDbMemoryService recalls past turns by keyword match on any MongoDB deployment, with no vector index and no embedding model.

Will my traces break after upgrading? They can. Telemetry now defaults to the OpenTelemetry GenAI 1.44 conventions and removes the call_llm span. Set ADK_TELEMETRY_SCHEMA_VERSION_OPT_IN=otel_semconv_1_36 to keep the old schema.

Are the MongoDB services production-ready? The release notes label both the session and memory services experimental, so treat them as such even though 12 other features moved to stable.

For more, see the adk-python tool hub, the Frameworks category, the earlier adk-python 2.12.0 launch and the Pydantic AI vs Google ADK comparison.