Spec2FaaS: using AI agents for autonomous serverless function development and deployment

The Function-as-a-Service (FaaS) serverless paradigm is increasingly popular, allowing users to focus on application development and abstracting away operational aspects (e.g., auto-scaling), which are automatically managed by the underlying platform. The advent of Large Language Models (LLMs) capable of code generation raises the question of whether FaaS can be extended further to abstract away the coding effort from users as well. Although recent studies have investigated the use of LLMs for function generation and deployment, they do not provide an end-to-end pipeline spanning from user requirements to function deployment. To fill this gap, this paper introduces a novel multi-agent system, Spec2FaaS, that autonomously turns natural language specifications into validated deployment on an open-source FaaS platform. Spec2FaaS decomposes development into specialized agents for code generation, test synthesis, execution, debugging, and deployment, enabling fully automated and self-correcting function development. Evaluated on the HumanEval+ benchmark, Spec2FaaS achieves competitive performance, demonstrating that fully autonomous, end-to-end serverless engineering is feasible and promising, with 90% of the functions successfully deployed and executed.