Playbooks · Coding tasks
Google ADK 2.14: A2A graduates, Teams joins in
Written November 8, 2026
ADK 2.14.0 graduates execution environments and A2A from experimental, adds a Teams runner and multi-run evals. What's stable now.
What graduates
Google published Agent Development Kit 2.14.0 on November 7. The release notes describe it as the release that graduates execution environments, A2A, evaluation, and core runtime modules from experimental, and adds multi-run evaluation, semantic tool argument matching, Microsoft Teams and OpenAI Decisions integrations, and faster long multi-turn workflows. In numbers, 26 auth, environment, app, CLI, and cache modules move to stable, alongside 16 toolset and integration features, plus the mature evaluation and user simulation classes.
Graduating changes what you can rely on. These modules are now stable APIs you can build on without experimental warnings, and the 16 promoted toolset features are always available: the feature flags that used to block them no longer do anything, with no error or warning. A2A is the one to note. It shipped as experimental before this release and graduates to stable now, with adk generate_agent_card building A2A agent cards from the command line. If you held off on A2A because it was experimental, that reason is gone. If you already built on it, the upgrade is mostly the removal of the warning, not a rewrite.
What is new
The headline addition is TeamsRunner, which connects an ADK agent to Microsoft Teams chats and channels, putting your agents where a team already talks. Next to it, MultiRunConfig runs each evaluation case multiple times concurrently, through adk eval --num_runs and a --parallelism flag, combining scores or pass rates across runs. That is a real change in how evals work: one pass of a case tells you what happened once, and a flaky agent can pass once by luck. Running each case several times and combining the results gives a steadier verdict, which is what you want before trusting an eval to gate a deploy.
ToolArgsMatcher compares tool call arguments in evaluations and environment simulation using exact, case-insensitive, numeric-tolerance, or LLM-backed semantic checks. The LLM-backed option matters for arguments where two wordings mean the same thing and an exact match would fail a correct answer. The release also adds OpenAIDecisionsClassifier, which classifies workflow and tool-gating questions with the OpenAI Decisions API, and vector-search retrieval with a time-to-live for MongoDbMemoryService.
The performance claim, and its caveats
Google says long multi-turn workflows run faster and use less memory, and gives one measurement: a 30-node workflow's 200-turn run is 23 percent faster with 9 percent lower peak memory. The release notes publish no baseline hardware, no second benchmark, and no comparison against a named prior ADK version. That makes the figure a single-configuration claim, not a general speedup. A 30-node, 200-turn workflow is a long, chatty run; a short workflow with few turns may see nothing like 23 percent. Treat the number as a reason to measure your own longest workflow after upgrading, not as a promised gain.
Breaking changes to check before upgrading
Three breaking changes ride along. Enum tool arguments now arrive as enum members instead of the raw strings the model sent, so code that compares an argument to a string, uses it as a dictionary key, or matches on it will break; compare against members, use .value, or switch to StrEnum. VertexAiSessionService.create_session now accepts only session config options, and other keys raise ValueError. And RunSkillScriptTool no longer takes a free-form command in environment mode: with a SkillToolset environment, the model passes args, short_options, and positional_args instead, scripts run from the skill directory rather than the environment working directory, and only .py, .sh, and .bash scripts are supported. Calls, eval sets, or prompts that still pass command now return INVALID_ARGUMENTS. The reason is a security hole: the old shape let a prompt-injected model run arbitrary shell commands. If your agents use skill scripts, update the calls before upgrading, or the upgrade will break them in a way that looks like the model misbehaving.
Limits
- This page is a read of the ADK 2.14.0 release notes. No ADK agent was built or run on this version.
- The 23 percent faster and 9 percent lower peak memory figures describe one 30-node, 200-turn configuration in Google's measurement. No baseline hardware, no named prior version, and no second benchmark are published.
- A2A graduates in this release; it is not new. It existed as experimental before 2.14.0.
- The release notes carry a November 4 heading date, while the release itself was published on GitHub on November 7.
Related playbooks
- Google's AQuA watches for silent agent failures
- Google Jules ends Nov 11: export your work
- Google put a TPU in orbit and got Gemma 4 to answer
- Antigravity's remote agents run in Google's cloud
Sources
- Release v2.14.0 — google/adk-python — the graduation list, TeamsRunner, MultiRunConfig, ToolArgsMatcher, the 23 percent and 9 percent figures, and the three breaking changes
Back to: Coding playbooks hub · Also: OpenAI's Agents API: the harness is now hosted · GitHub Agentic Workflows are now GA · Google Jules ends Nov 11: export your work