arXiv:2604.00137 introduces OpenTools, a community-driven toolbox that standardizes tool schemas, supplies plug-and-play wrappers, and applies automated test suites with continuous monitoring to address reliability in tool-integrated LLMs (https://arxiv.org/abs/2604.00137). The work states that failures arise from both tool-use accuracy and intrinsic tool accuracy, with prior emphasis mainly on the former. Experiments reported 6%-22% relative gains over an existing toolbox across agent architectures on downstream tasks.

ReAct (https://arxiv.org/abs/2210.03629) established prompting methods for reasoning and acting with tools yet did not detail ongoing tool-quality monitoring or community contribution protocols. Toolformer (https://arxiv.org/abs/2302.04761) demonstrated LLMs learning to call external tools but similarly omitted collective maintenance mechanisms and public test-case evolution described in the new framework.

The released public web demo permits users to run agents, contribute test cases, and generate evolving reliability reports as tools change, per the primary source. OpenTools supplies the core framework, initial tool set, evaluation pipelines, and contribution protocol, improving end-to-end reproducibility.