Low-resource languages remain difficult for large language models because suitable training data, evaluation resources, and expert supervision are limited. Adapting a model to a new language usually requires a human trainer to inspect its errors, identify missing knowledge, prepare additional data, choose a training method, and evaluate whether the model has improved. This process is effective, but it is slow, costly, and difficult to repeat for many languages. This project studies whether an AI agent can take over part of this process through recursive self-improvement. We define recursive self-improvement as a controlled sequence of training cycles. In each cycle, the agent evaluates the current model, identifies important weaknesses, selects or generates targeted training examples, chooses a training intervention, runs the training code, and verifies the resulting model. The results from one cycle are then used to plan the next cycle. The agent will operate within a fixed computational budget, a restricted set of tools, and predefined evaluation criteria. The main research question is how close such an agent can come to a human trainer when adapting a language model to a previously underrepresented language. We will compare three settings under matched data and compute budgets: a fixed training pipeline, an iterative pipeline managed by human trainers, and an iterative pipeline managed by an AI agent. Human trainers and the agent will receive the same model outputs, error reports, training tools, and limits on the number of experiments. This design will allow us to compare final model quality and the quality of the training decisions made along the way. The experiments will focus on several low-resource languages and open language models of different sizes. We will use datasets and evaluation protocols from WMT shared tasks to study machine translation, robustness, domain transfer, and learning with limited parallel data. We will also use multilingual tasks from SemEval shared tasks, including semantic similarity, sentiment and stance analysis, information extraction, and language understanding in culturally specific settings. Together, these tasks will test whether the agent can diagnose and address weaknesses beyond translation alone. We will measure language understanding, generation, translation, grammatical accuracy, robustness, and transfer across languages. We will also compare improvement per training example, improvement per GPU hour, the number of useful interventions, and the ability to recover from unsuccessful experiments. A further question is whether the agent can transfer a successful learning strategy from one language or task to another. The project will provide a direct measurement of the gap between human-guided and agent-guided model improvement. It will identify which parts of language-model adaptation can be automated reliably and which still require human expertise. The expected outcome is a practical and reproducible framework for studying recursive self-improvement under limited data and computing resources, rather than an unrestricted system that modifies itself indefinitely. The results may help reduce the cost of developing language technologies for communities that are poorly represented in current foundation models.