Google Researchers Open-Source EnvHarness for Adaptive Agent Training
Google researchers have open-sourced EnvHarness, a framework that adapts training environments to an agent’s weaknesses without modifying the agent, underlying environment, or evaluation system. It wraps existing environments through standard reset() and step() interfaces and includes EnvRigger to identify weaknesses and generate targeted changes. The researchers reported gains of 2.9–3.7 percentage points across WebArena, SWE-bench Verified, and ALFWorld for several tested models.
EnvHarness offers a domain-independent way to vary training conditions for self-evolving agents while preserving existing environments and evaluators, potentially addressing the limited learning opportunities posed by fixed benchmarks.

谷歌研究人员开源 EnvHarness",这是一个能够根据智能体的短板动态调整智能体训练环境的框架,且无需修改智能体、环境或评估系统。研究人员表示,这种方法既能提高学习效果和效率,又能帮助智能体克服自身缺陷。
研究人员表示,自我进化型智能体通常通过重写提示词和获取新技能来改进自身,但很少会对其运行的环境进行调整。这带来了一个关键问题:
一个在固定基准上不断改进的智能体最终将无法再从该基准中学到东西。对于处于每个学习阶段的每个学习者而言,初始状态、规则和反馈都是相同的,因此一旦智能体能够解决这些任务,环境就再也无法提供新的知识。
研究人员解释说,环境生成成本高昂,需要特定领域的流水线,并且往往需要依赖昂贵或不可靠的验证器。即便投入大量资源,生成出来的环境仍然是静态的。为了让工程师更容易在自适应环境中训练智能体,他们开发了 EnvHarness"。
#Google#EnvHarness#EnvRigger#WebArena#SWE-bench Verified#ALFWorld#Gemini 3.1 Flash-Lite#Qwen3.6 27B#Gemini 3.5 Flash#Claude Sonnet 4.6