ADK 2.14.0 graduates execution environments, A2A, and evaluation from experimental

The ADK v2.14.0 release, published on November 7, graduates execution environments, A2A, evaluation, and core runtime modules from experimental to stable. It adds multi-run evaluation, semantic tool argument matching, and integrations for Microsoft Teams and OpenAI Decisions, while cutting long multi-turn workflow runtimes by 23%.

Execution environments and sandboxes reach stability

BaseEnvironment, LocalEnvironment, and platform-specific environments for Daytona and E2B are now stable APIs. The update introduces streaming file reads so that `ReadFileTool` can grab line ranges without loading entire files into memory. For large-scale operations, `AgentEngineEnvironment` reads and writes large files in chunks and reports commands running longer than 570 seconds as timed out before hitting the service deadline.

`ContainerCodeExecutor` now supports input and output files. A related security change closes a hole where a prompt-injected model could run arbitrary shell commands: `RunSkillScriptTool` no longer accepts free-form shell commands in environment mode. The model must now pass structured arguments, and only `.py`, `.sh`, and `.bash` scripts are allowed.

Evaluation gets multi-run and semantic matching

Evaluators can now run each test case multiple times concurrently using `MultiRunConfig` and `--parallelism`, combining scores or pass rates across runs for steadier verdicts. The new ToolArgsMatcher compares tool call arguments using exact, case-insensitive, numeric-tolerance, or LLM-backed semantic checks. `AgentEvaluator` can evaluate agents built at runtime, and `--eval_storage_uri` lets teams store evaluation files outside a read-only agents directory.

New integrations and workflow performance

A stable A2A integration lets users build agent cards from the CLI with `adk generate_agent_card`. The release adds a TeamsRunner to connect ADK agents to Microsoft Teams chats and channels, and an OpenAIDecisionsClassifier that classifies workflow and tool-gating questions using the OpenAI Decisions API. MongoDbMemoryService gains vector search retrieval and TTL expiration.

On the performance side, a 30-node workflow's 200-turn run is now 23% faster with 9% lower peak memory. Token-threshold compaction now counts tool outputs since the last model call, triggering earlier for an unchanged threshold.

Breaking and behavior changes to note

Enum tool arguments now arrive as members rather than raw strings. A plain `Enum` argument will no longer work with string comparisons, dict lookups, or `json.dumps` - teams should compare against members, use `.value`, or switch to `StrEnum`. `VertexAiSessionService.create_session` now accepts only session config options; extra keyword arguments raise `ValueError`.

`DatabaseSessionService` creates new PostgreSQL JSON columns as `json` instead of `jsonb`. Existing `jsonb` tables keep working, but queries relying on jsonb-only operators or GIN indexes won't function on new tables. Sixteen toolset and integration features promoted to stable - including Bigtable, Computer Use, and MCP agent server - are now always available and can no longer be blocked by feature flags.

Why this matters for IT and development professionals

For teams building agent workflows, the 23% speed improvement on complex multi-turn runs and the lower memory footprint translate directly to reduced infrastructure costs. The shift to structured tool arguments and the closure of the shell command injection vector in `RunSkillScriptTool` address a concrete security risk for anyone running models that process untrusted input. If your team relies on PostgreSQL session storage, plan for the `json` versus `jsonb` change before spinning up new tables - existing deployments won't break, but new schemas will lack jsonb indexing unless you migrate manually. Professionals looking to build these skills can explore AI for Developers Courses that cover agent architecture and deployment patterns.