Google ADK Python v2.10.0 (GitHub 2026-09-25; changelog 2026-09-24) adds experimental skill lifecycle controls (EPHEMERAL, active-skill caps, unload_skill via ADK_ENABLE_SKILL_LIFECYCLE=1), a MongoDB toolset with vector/hybrid search, eval efficiency metrics for duration/tokens/model calls, and better OpenAI reasoning param adaptation plus reasoning-token reporting. PyPI google-adk 2.10.0 is live.

Key Takeaways

  • ✓Shipped: GitHub v2.10.0 + PyPI google-adk 2.10.0; docs at adk.dev/get-started
  • ✓Skills (experimental): enable ADK_ENABLE_SKILL_LIFECYCLE=1 for EPHEMERAL one-turn lifecycle, active-skill caps, and opt-in unload_skill
  • ✓Data: MongoDB toolset with vector and hybrid search inside agent flows
  • ✓Eval: duration efficiency metric plus token and model-call counts
  • ✓Models: OpenAI reasoning param adaptation and reasoning-token reporting; use effort instead of thinking_config on OpenAIResponsesLlm; adk create adds gemini-3.8-flash
🔗

Project Links & Resources

Direct Access
Direct access to official project resources and documentation
🐙
ADK Python v2.10.0 Release
github.com/google/adk-python/releases/tag/v2.10.0
↗
📖
google-adk 2.10.0 on PyPI
pypi.org/project/google-adk/2.10.0/
↗
📖
ADK Docs (adk.dev get-started)
adk.dev/get-started/
↗
📖
ADK Docs mirror (github.io)
google.github.io/adk-docs/
↗
🐙
ADK CHANGELOG
github.com/google/adk-python/blob/main/CHANGELOG.md
↗
🔬

In-Depth Technical Analysis

Background

Multi-agent stacks need skill/tool catalogs, vector retrieval, and cost-aware evaluation in one Python runtime. Pain points: always-on skills blow context, DB search needs another client, evals ignore duration/tokens/calls, and OpenAI reasoning params/token accounting disagree. Google ADK Python v2.10.0 (released 2026-09-25; changelog dated 2026-09-24) targets those frictions.

Architecture

Experimental skill lifecycle via ADK_ENABLE_SKILL_LIFECYCLE=1: EPHEMERAL (one turn), active-skill caps, opt-in unload_skill (dropped from later requests), plus SkillDiscoveryMode. New MongoDB toolset for in-flow vector/hybrid search. OpenAI moves to google.adk.integrations.openai (google-adk[openai]); reasoning request adaptation and reasoning-token reporting; OpenAIResponsesLlm ignores thinking_config—use OpenAIGenerateContentConfig.effort. BigQuery protected write only allows non-SELECT when dry-run lands in the session anonymous dataset; protected-mode sessions stay in-process memory.

Benchmarks

No official SWE/AgentBench numbers. Eval now tracks duration, tokens, and model-call counts; empty eval sets raise ValueError instead of silently passing. Compare before/after skill lifecycle on your own suites.

Getting started

pip install -U google-adk==2.10.0 ([openai] extra optional). Docs: adk.dev/get-started and google.github.io/adk-docs. Enable skills: export ADK_ENABLE_SKILL_LIFECYCLE=1. adk create adds gemini-3.8-flash. Migrate legacy live-audio imports under google.adk.flows.llm_flows to google.adk.live.