Illustrative image: Python code. Google ADK is a Python framework for building AI agents. Image: Thraea19 / Wikimedia Commons, CC BY-SA 4.0
📌 Key Highlights
- Google released ADK 2.14.0 on 7 November 2026: 26 modules and 16 toolset features graduate from experimental to stable, along with the A2A integration and the evaluation classes.
- New TeamsRunner connects an ADK agent to Microsoft Teams chats and channels, so an agent can work where teams already talk.
- MultiRunConfig (
adk eval --num_runs) runs each eval case several times at once and combines the scores, so flaky single runs stop misleading you. - ToolArgsMatcher checks tool-call arguments exactly, case-insensitively, with numeric tolerance, or semantically with an LLM.
- Three breaking changes ship in the same release: enum tool arguments, Vertex AI session creation, and skill-script execution — all three need code changes in existing agents.
Google ADK 2.14 new features explained: what actually changed
Google ADK — the Agent Development Kit — is Google's open-source Python framework for building AI agents: the programs that plan, call tools, remember context and carry multi-step work to completion. Version 2.14.0 is a maturity release rather than a new-product launch. Its headline is not one feature but a status change: after months in experimental, the parts developers actually build on — execution environments, agent-to-agent communication, evaluation and the core runtime — are now stable, which means Google is telling teams they can build production agents on them without experimental warnings.
The release notes describe it plainly: the update "graduates execution environments, A2A, evaluation, and core runtime modules from experimental, adds multi-run evaluation and semantic tool argument matching, adds Microsoft Teams and OpenAI Decisions integrations, and makes long multi-turn workflows faster." Here is what each of those pieces means in practice.
| What changed | What it does | Why it matters |
|---|---|---|
| 26 modules go stable | Auth, environment, app, CLI and cache modules leave experimental status | Production agents can depend on them without experimental warnings |
| 16 toolset features go stable | Bigtable, Computer Use, Daytona and E2B environments, GCS, MCP agent server, Pub/Sub, Spanner and more are always available | No more feature flags to enable them; ADK_DISABLE_<FEATURE> no longer blocks them |
| TeamsRunner | Connects an ADK agent to Microsoft Teams chats and channels | Agents meet users inside Teams instead of a separate app |
| MultiRunConfig | adk eval --num_runs plus --parallelism runs each eval case several times concurrently and combines scores or pass rates | Steadier eval verdicts; one lucky or unlucky run cannot hide a regression |
| ToolArgsMatcher | Compares tool-call arguments exactly, case-insensitively, with numeric tolerance, or with LLM-backed semantic checks | Evals can accept "₹1,000" and "1000" as the same answer, or judge meaning instead of spelling |
| A2A goes stable | The Agent2Agent integration is stable; adk generate_agent_card builds an A2A agent card from the CLI | Agents from different frameworks can discover and talk to each other through a standard card |
| OpenAIDecisionsClassifier | Classifies workflow and tool-gating questions with the OpenAI Decisions API | Lets an agent ask a specialised model "should this tool run?" before acting |
| MongoDbMemoryService upgrades | Vector search retrieval (retrieval_mode="vector" or "auto") and TTL expiry (memory_ttl_days) | Agent memory can now search by meaning and forget old entries automatically |
| Execution environments stable | BaseEnvironment, LocalEnvironment, EnvironmentToolset, DaytonaEnvironment and E2BEnvironment are stable; files stream in chunks instead of loading whole files into memory | Agents can work with large files without running out of memory; long commands (over 570 seconds) are now reported as timed out instead of hanging |
| Faster long workflows | Google says a 30-node workflow's 200-turn run is 23% faster with 9% lower peak memory | Long agent conversations cost less time and memory — but read the caveat below before quoting the number |
Two smaller additions are worth knowing. The CLI gains adk run --reload_agents for hot-reloading agents while you develop, and session rewind is now available as a REST endpoint, so an application can roll an agent's session back without going through the Python API.
Microsoft Teams is the headline integration
Microsoft Teams logo. ADK 2.14's TeamsRunner connects an ADK agent to Teams chats and channels. Image: Sosalmediahelp / Wikimedia Commons, CC BY-SA 4.0
Of all the new integrations, TeamsRunner is the one most teams will actually use first. It connects an ADK agent to Microsoft Teams chats and channels, which means an agent built with Google's framework can answer questions, run workflows and report results inside the chat tool where the team already works — no separate agent app to open, no new interface to learn. For companies that run on Teams, that removes the biggest adoption hurdle for an internal agent: getting people to go somewhere else to use it.
The same release quietly adds an OpenAI Decisions classifier, which routes workflow and tool-gating questions through OpenAI's Decisions API — a reminder that ADK is not a Google-only walled garden: it will happily call another provider's model when that model is the right tool for one specific decision.
Three breaking changes to fix before you upgrade
ADK 2.14 is not a drop-in upgrade for every project. Three breaking changes ship in this release, and each one can break existing agent code silently or loudly. Check all three before you run the update.
| Breaking change | What breaks | The fix |
|---|---|---|
| Enum tool arguments arrive as members | FunctionTool parameters annotated with an Enum now receive enum members instead of the raw strings the model sent, so arg == "value", string-keyed dict lookups and string methods stop working | Compare against members, use .value, or switch to StrEnum |
| VertexAiSessionService.create_session tightens its arguments | Extra keyword arguments are now limited to display_name, ttl, expire_time and labels; anything else raises ValueError (HTTP 400 through adk api_server) | Pass state with state= and IDs with session_id= instead of stuffing them into session options |
| RunSkillScriptTool drops the free-form command | In environment mode the tool no longer takes a command string; calls that pass one now return INVALID_ARGUMENTS | Pass args, short_options and positional_args instead, and keep scripts in the skill directory as .py, .sh or .bash files |
The third change is the most important of the three, and it is a security fix, not just a tidy-up. A free-form shell command parameter meant a prompt-injected model could get the agent to run arbitrary shell commands — the classic way an agent escapes its sandbox. Restricting skill scripts to validated paths and arguments from a designated directory closes that hole. If your agents use skill scripts, update the calls: this is the one breaking change you should not postpone.
One more upgrade warning from the release notes: FirestoreSessionService now stores event payloads as JSON strings, and sessions written by 2.14 cannot be read by 2.13.0 or earlier if you roll back. Test the upgrade on a copy of your data first, and do not roll back a live Firestore-backed deployment without a migration plan.
Why it matters for Indian developers
India has one of the world's largest communities of developers building AI agents, and ADK is one of the frameworks they reach for first — it is Python, it is open source, and it runs on Google Cloud's India regions as well as locally. A stable 2.14 matters here in three concrete ways. First, the Teams integration lands where Indian enterprises actually work: Microsoft Teams is standard issue in Indian IT services firms, banks and startups, so an agent that lives inside Teams is an agent employees will actually meet. Second, multi-run evaluation answers a real pain for teams shipping agents to production: a single eval run is noisy, and a noisy eval is how a bad agent ships. Running each case several times and combining the scores is a small change that makes "did we break the agent?" a question with a trustworthy answer. Third, the breaking changes are exactly the kind of thing that bites during a Monday-morning upgrade — the enum change in particular will break tool code that compares arguments to plain strings, a pattern every ADK tutorial teaches.
There is no India-specific announcement in this release — no India pricing, no India-only feature — and this post does not invent one. The India angle is that Indian developers are among the heaviest users of exactly the tools this release stabilises, and the upgrade checklist above is written for them.
Should you upgrade to ADK 2.14?
Yes — but test first. ✅ Verdict: upgrade, after a short checklist. If you build agents with ADK, 2.14 is the release the framework has been building toward: the stable modules, the A2A support and the evaluation tooling are all things you will want within months, and the Teams runner is genuinely useful. But do the upgrade in a branch, not on main: grep your tools for enum comparisons, check your Vertex AI session calls, and rewrite any skill-script calls that pass a command string. Then run your eval suite with adk eval --num_runs and let the multi-run scores tell you whether the upgrade changed any behaviour. If you are on 2.13 and your agents are in production, wait for a quiet window and keep a rollback plan that accounts for the Firestore one-way door. ⏳ Verdict: upgrade in a branch, ship after evals pass.
One number deserves a caution flag of its own. Google says a 30-node workflow's 200-turn run is 23% faster with 9% lower peak memory in 2.14 — but the release notes publish no baseline hardware, no second benchmark and no comparison against a named earlier ADK version, and both independent write-ups of the release flag exactly that. Treat 23% as Google's own measurement on its own workload, not as a promise about your agent. The honest way to use the number is as a reason to measure your own workflows before and after the upgrade, not as a reason to skip the testing.
FAQs
What is Google ADK 2.14?
Google ADK 2.14.0 is the 7 November 2026 release of Google's Agent Development Kit, the open-source Python framework for building AI agents. It graduates 26 modules and 16 toolset features from experimental to stable, adds Microsoft Teams and OpenAI Decisions integrations, and brings multi-run evaluation and semantic tool-argument matching.
What is TeamsRunner in ADK 2.14?
TeamsRunner is a new ADK 2.14 component that connects an ADK agent to Microsoft Teams chats and channels. It lets an agent answer questions and run workflows inside Teams, where the team already works, instead of requiring a separate app. It arrives with a new teams install extra.
Is ADK 2.14 a breaking-change release?
Yes, it carries three breaking changes: enum tool arguments now arrive as enum members instead of raw strings, VertexAiSessionService.create_session rejects extra keyword arguments, and RunSkillScriptTool no longer accepts a free-form shell command. The last one is a security fix that closes a prompt-injection hole. Existing agents need code changes before upgrading.
Should I upgrade to Google ADK 2.14 now?
Yes, but test first. Upgrade in a branch with pip install -U google-adk, fix the three breaking changes, and run your eval suite with adk eval --num_runs to confirm behaviour is unchanged. Production deployments on Firestore should plan around the fact that 2.14-written sessions cannot be read by 2.13.0 or earlier after a rollback.
Takeaway
ADK 2.14 is the release where Google's agent framework grows up: 42 modules and features go stable, agents move into Teams, and evaluations get the repeat runs they always needed. The price of that maturity is three breaking changes — one of them a genuine security fix — so upgrade in a branch, fix the enum and skill-script calls, and let multi-run evals prove nothing changed before you ship.
If you build agents with Google's framework, the next useful read is Meta's Astryx design system — another developer-tool release from the same week, and a good sense of where agent tooling is heading. For the other side of the agent story, see coding agents you can talk to and how Claude fixes its own notes between sessions.
Sources
- google/adk-python v2.14.0 release notes (GitHub) — primary source: full changelog, highlights, breaking changes and behaviour changes
- Google ADK 2.14 Adds Teams Integration and Multi-Run Evals (Emergent) — independent write-up, flags the missing performance baseline
- Google ADK 2.14.0 Graduates Core Modules, Adds Teams Runner (AI Weekly) — independent write-up, confirms the 42 stable components and the three breaking changes
Comments
Post a Comment