Abstract
As large language models (LLMs) become increasingly common in handling complex tasks that require external tools and services, the various issues that arise when calling tools have become a major factor affecting system stability and usability. This article addresses the situation where LLMs fail to learn or botch the use of tools, and proposes a practical solution for identifying and resolving such issues. The key actions taken include: First, we established a framework to categorize all cases of tool learning failures and developed a tester capable of intentionally inducing faults; Then, we proposed a runtime recovery framework called Monitor-Classifier-Recovery (MCR), which includes simple rules, a fault classifier capable of learning, and a strategy library that considers costs; Then, we introduced the MAGRPO method, which involves multiple agents optimizing strategies through a process similar to a multi-player rock-paperscissors game. During training, agents collaborate and can compensate for each other's shortcomings. We tested this approach through a combination of writing and programming tasks, such as format errors, timeouts, returning garbled characters, and sequential errors in planning. Using MCR alone significantly accelerated problem resolution and reduced human intervention. Combining MCR and MAGRPO boosted the success rate for complex tasks from 42 % to 86.7 %, with a problem resolution rate of 92.6 %, while also optimizing API costs and speed. The article also discusses whether recovery strategies can be combined like building blocks, how to balance costs and benefits, how to model API behavior in the real world, and ethical reviewthese are engineering and research challenges. This approach provides a clear path for LLM systems to become more reliable, easier to understand, and efficient when using tools in practice.