Google ADK Python 2.10 adds one-turn ephemeral skills and cost metrics, and empty eval sets now fail

GitHub (google/adk-python)·low signalXBlueskyLinkedInCopy link

ADK 2.10.0, released 25 September, adds an experimental skill lifecycle (ADK_ENABLE_SKILL_LIFECYCLE=1) with an EPHEMERAL mode that lasts one turn, active-skill limits and an opt-in unload_skill tool, plus a MongoDB vector and hybrid search toolset. Evals now report duration, token use and model-call counts. Breaking behavior: AgentEvaluator.evaluate raises ValueError when no eval cases run instead of passing, ${var} in instructions is no longer filled from state, and BigQuery protected write mode rejects CALL, EXPORT DATA and multi-statement scripts. A CI job that passed on an empty eval set will fail after upgrading.