Google ADK (Python) v2.14.0

Summary

Google's Agent Development Kit for Python 2.14.0 was published on GitHub on 2026-11-07 (its changelog entry is dated 2026-11-04). Execution environments and the A2A integration are now stable. Evals can repeat each case several times, and tool-argument checks can match by meaning. It has one breaking change for enum tool arguments.

Key points

• Stable execution environments: BaseEnvironment, LocalEnvironment, EnvironmentToolset, DaytonaEnvironment, and E2BEnvironment leave experimental. ReadFileTool can read a line range instead of the whole file.
• A2A stable: a new adk generate_agent_card command builds agent cards.
• Multi-run evals: adk eval --num_runs N --parallelism M runs each case several times and combines the results, which helps with flaky agent behaviour.
• ToolArgsMatcher: compares tool arguments by exact, case-insensitive, numeric-tolerance, or LLM-judged semantic match.
• Integrations: TeamsRunner (Microsoft Teams), OpenAIDecisionsClassifier (OpenAI Decisions API), and MongoDbMemoryService with vector search and TTL expiry.
• Performance: about 23% faster and about 9% lower peak memory on a 30-node, 200-turn workflow (Google's figures).
• Breaking: enum tool arguments now arrive as enum members, not strings; code that compares them with strings needs updating.
• Upgrade: pip install -U google-adk==2.14.0.

Why it matters

Repeated eval runs and semantic argument matching are practical answers to measuring agents that do not behave the same way twice.