Stable execution environments and tool safety
Google released Agent Development Kit version 2.14.0, marking the graduation of its execution environments out of experimental status. Sandboxes such as LocalEnvironment, DaytonaEnvironment, and E2BEnvironment are now marked stable. To avoid out-of-memory errors on large targets, BaseEnvironment adds chunked file streaming so ReadFileTool can inspect line ranges without loading entire files into memory.
The release also updates security around skill script tools. In environment mode, RunSkillScriptTool no longer accepts free-form shell commands. Agents must pass explicit arguments, short options, and positional parameters from skill directories. Execution is restricted to Python and shell scripts, blocking prompt injection attacks that attempt to execute arbitrary shell payloads.
Evaluation enhancements and runtime interfaces
For testing agent systems, ADK adds MultiRunConfig and parallel CLI flags to run evaluation cases concurrently and aggregate scores across multiple runs. In addition, ToolArgsMatcher lets developers evaluate tool call arguments during tests and environment simulations using exact matches, case-insensitive checks, numeric tolerance, or language-model semantic comparisons.
Agent-to-agent communication modules also graduate to stable status, accompanied by command line tools to generate standard agent cards. New runtime components include TeamsRunner for Microsoft Teams channels and OpenAIDecisionsClassifier for gating workflow steps using the OpenAI Decisions API.
Breaking changes accompany the release. Function tools with Enum parameter annotations now receive actual enum members instead of raw strings, breaking plain string comparisons unless updated to StrEnum. Google also reports multi-turn workflow runtime speedups, claiming a 23 percent faster completion time and 9 percent lower peak memory in internal 30-node benchmarks.
Our assessment
Teams building production workflows with ADK should audit custom tool definitions that rely on string-matched Enum parameters, as passing members directly will break existing string comparisons unless switched to StrEnum. Security teams using SkillToolset should verify that automated prompts and evaluation suites no longer emit raw shell command strings, ensuring all skill calls migrate to explicit script paths and argument lists.
Source: Google original publication.
Read the original material
Google (opens in a new tab)The article date belongs to Agentic Horizon. The source publication date is listed separately. Vendor claims remain attributed; we have not independently tested the reported capability.
