Tool Suggestions
Google ADK 2.14.0 Makes Agent-to-Agent, Environments and Evaluation Stable
Google's Agent Development Kit, ADK, reached version 2.14.0 this week: the release notes carry the date November 4, and the GitHub tag was published on November 7. The headline is a graduation. Execution environments, the agent-to-agent integration, evaluation and the core runtime all move from experimental to stable, which means developers can build on them without experimental warnings. ADK is an open-source, code-first Python toolkit for building, evaluating and deploying AI agents, model-agnostic and deployment-agnostic and optimised for Gemini, under the Apache 2.0 licence. It installs with pip install google-adk on Python 3.11 and up, and also exists for Java, Kotlin, Go and TypeScript.
What graduated to stable
The stable set covers execution environments — BaseEnvironment, LocalEnvironment, EnvironmentToolset, DaytonaEnvironment and E2BEnvironment — plus the A2A integration, evaluation and the core runtime. Behind the headline are 26 auth, environment, app, CLI and cache modules, 16 toolset and integration features, and mature evaluation and user-simulation classes. The release notes sum it up directly: “This release 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.”
Agent-to-agent and evaluation
On the A2A side, a new command, adk generate_agent_card, builds an A2A agent card from the command line, and task cancellation in A2A now propagates to active runner invocations, so cancelling a task actually stops the work. Evaluation gets the two features that matter most for anyone testing agents seriously. MultiRunConfig, used through adk eval --num_runs with a --parallelism flag, runs each test case several times concurrently and combines the scores and pass rates, which is how you tell a genuinely better agent from a lucky run. And AgentEvaluator now accepts a BaseNode or an App built at runtime, with an --eval_storage_uri option for where results go.
The more interesting evaluation addition is ToolArgsMatcher, which compares the arguments of a tool call rather than just checking that the right tool was called. It supports exact matching, case-insensitive matching, numeric tolerance and LLM-backed semantic checks, and it is also used in Environment Simulation, where an agent is tested against a simulated environment instead of the real one.
New integrations
TeamsRunner connects an ADK agent to Microsoft Teams chats and channels, arriving as a new teams install extra — an agent built with Google's toolkit can now live where a company's people already talk. OpenAIDecisionsClassifier adds an integration with the OpenAI Decisions API. And MongoDbMemoryService gains vector-search retrieval, with retrieval_mode set to vector or auto, plus time-to-live expiry for stored memories through memory_ttl_days, so an agent's memory can be searched by meaning and can forget on a schedule.
Environments and long runs
Execution environments get a set of practical improvements. BaseEnvironment.read_file_stream adds chunked file streaming, AgentEngineEnvironment handles chunked large files and reports commands running over 570 seconds as timed out, and ContainerCodeExecutor gains input and output files. RunSkillScriptTool now runs skill scripts from validated paths and arguments instead of a shell string. There is a --reload_agents hot-reload flag, MCP toolset callers can receive notifications from MCP servers, an experimental FsspecArtifactService and BigQuery skills through get_bigquery_skill, an e2b floor of 2.29.5, and a new fsspec extra.
The performance claim
Google says “a 30-node workflow's 200-turn run is 23% faster with 9% lower peak memory” in this release. That is Google's own claim, and it should be read as one: no hardware, no baseline ADK version and no methodology are published with it, and the cited commit, which scopes event indexing in the workflow replay manager, repeats the figure without any benchmark setup. Long multi-turn workflows may well be faster; the number itself cannot be checked from what Google has published.
Breaking changes to check before upgrading
Several changes in 2.14.0 can break existing code. FunctionTool parameters annotated with an Enum now receive enum members instead of raw strings, so a comparison like arg == “value” will stop working. RunSkillScriptTool no longer takes a command argument in environment mode — the model passes args, short_options and positional_args instead, and only .py, .sh and .bash files are allowed; the notes say this “closes a hole that let a prompt-injected model run arbitrary shell commands”, which is a security fix worth taking even if it means changing your code. VertexAiSessionService.create_session now accepts only session configuration keys and raises a ValueError for anything else. Sessions written by FirestoreSessionService in 2.14.0 cannot be read by 2.13.0 or earlier, so a downgrade after upgrading will lose session data. The 16 promoted features are always available now, and ADK_DISABLE_<FEATURE> no longer blocks them. And DatabaseSessionService creates new PostgreSQL JSON columns as json instead of jsonb, though existing tables are not migrated. None of these is a reason to stay on 2.13.0, but all of them are reasons to read the release notes before upgrading a production agent.