adk-python 2.14.0 makes sandboxes and A2A stable

Google published adk-python v2.14.0 on GitHub on November 7, 2026 (the changelog heading is dated November 4). Like 2.13.0, it is mostly a stabilization release: execution environments, the A2A integration and the mature evaluation classes leave experimental status. It also adds multi-run evaluation, a Microsoft Teams runner and vector retrieval for MongoDB memory. Three breaking changes are listed, and one of them closes a prompt-injection hole in skill scripts.

Release facts

What changed

• Sandboxes are stable. BaseEnvironment, LocalEnvironment, EnvironmentToolset, DaytonaEnvironment and E2BEnvironment graduate from experimental. A new read_file_stream reads files in chunks, so ReadFileTool can return a line range without loading a whole file.
• A2A is stable. The A2A integration leaves experimental, and adk generate_agent_card builds an A2A agent card from the CLI.
• Multi-run evaluation. adk eval --num_runs with --parallelism runs each eval case several times concurrently and combines scores or pass rates. That matters for agents whose output varies from run to run.
• Smarter tool-argument checks. ToolArgsMatcher compares tool-call arguments with exact, case-insensitive, numeric-tolerance or LLM-backed semantic matching.
• New integrations. TeamsRunner connects an agent to Microsoft Teams chats and channels, and OpenAIDecisionsClassifier uses the OpenAI Decisions API for workflow and tool-gating questions.
• MongoDB memory gets vectors. MongoDbMemoryService, which in 2.13.0 matched past turns by keyword only, adds retrieval_mode="vector" or "auto" plus TTL expiry through memory_ttl_days.
• Performance. The notes report that a 200-turn run of a 30-node workflow is 23% faster with 9% lower peak memory. That is Google's own measurement.

Breaking changes to check before upgrading

1. Enum arguments arrive as enum members. FunctionTool parameters typed Enum, Optional[Enum] or list[Enum] now receive members, not the raw strings the model sent. Code like arg == "value" or a string-keyed dict lookup stops matching. Compare against members, use .value, or switch to StrEnum.
2. Stricter Vertex AI session options. VertexAiSessionService.create_session only accepts display_name, ttl, expire_time and labels as extra options. Keys such as http_options or session_state now raise ValueError; pass state= and session_id= instead.
3. No free-form shell in skill scripts. In environment mode, RunSkillScriptTool takes args, short_options and positional_args instead of a command string, and only runs .py, .sh and .bash scripts. The notes say this closes a hole that let a prompt-injected model run arbitrary shell commands. Eval sets or prompts that still pass command get INVALID_ARGUMENTS.

The behavior changes deserve a look too. FirestoreSessionService now writes event payloads as JSON strings, and the notes warn that sessions written by 2.14.0 cannot be read by 2.13.0 or earlier after a rollback. New PostgreSQL tables from DatabaseSessionService use json instead of jsonb, so outside queries that rely on jsonb operators or GIN indexes will not work on them.

Who should upgrade now

Teams that give ADK agents a sandbox, through E2B, Daytona or a local environment, get the most from this release: the APIs are stable and the skill-script change removes a real injection path. Teams that run evals in CI gain --num_runs for flaky agents. If you use Firestore sessions, upgrade on a branch first, because rolling back after new sessions are written is not supported.

Questions

Is adk-python 2.14.0 safe to roll back? Not fully if you use Firestore sessions. The release notes say sessions written by 2.14.0 cannot be read by 2.13.0 or earlier.

Does ADK now support E2B and Daytona sandboxes officially? Yes. E2BEnvironment and DaytonaEnvironment are listed as stable in this release, and the e2b dependency floor rises to 2.29.5.

Why did my enum comparisons stop working after upgrading? FunctionTool now passes enum members instead of strings. Compare against the member or use .value.

For more, see the adk-python tool hub, the earlier adk-python 2.13.0 launch, the new Google ADK alternatives guide and the Frameworks category.