Every network device has a configuration file containing the current settings and operational functionality such as IP addresses, running routing processes, and filtering rules. When new functionality is requested, the network administrator updates the configuration file by adding new commands. This process can be automated using an LLM model that understands the configuration language and can generate the relevant configuration commands. This paper presents a multi-agent LLM system that generates network configurations. Our approach uses a sequence of LLM-based agents to decompose the original configuration task, expressed in natural language, into logical subtasks, which are then translated into configuration commands. The proposed LLM-based configuration generation process includes validation and suggests sanity tests to eliminate potential errors. We demonstrate our approach by generating Cisco IOS configuration files for multiple devices. We also propose a dataset of non-trivial reference configurations for evaluating generation accuracy.

The reliability of a network is a crucial requirement for systems such as IoT, client-server, or cloud-based solutions. Unfortunately, real networks cannot be assumed to be fault-free, especially when considering various hardware problems, performance issues, or malicious attacks. Testing networked systems should therefore include evaluating fault tolerance under various network conditions. The paper presents a doctoral research project on automated verification of networked systems using fault-attack injection using a derived model of network communication.