The heterogeneity of Internet of Things (IoT) devices remains a major barrier to scalable deployment, automated onboarding, and cross-platform interoperability. Device-specific protocols, software dependencies, telemetry formats, and runtime constraints often require manual configuration and ad-hoc integration logic. This paper proposes an agentic AI architecture for zero-touch IoT device onboarding, where a large language model is used as an autonomous configuration agent capable of generating device-side software, installing dependencies, validating runtime behavior, and recursively correcting failures until convergence is achieved. The proposed architecture introduces a closed-loop workflow that combines device inspection, AI-based runtime generation, remote deployment, telemetry validation, and self-correction. Interoperability is addressed through a normalized telemetry contract that abstracts device-specific behavior into a common telemetry representation. We evaluate the approach on a real IoT device by benchmarking multiple AI models under clean installation conditions in two campaigns: the original pipeline and a second pipeline assisted by a locally trained semantic context optimizer. The comparison covers convergence time, correction cycles, recursive iterations, and prompt tokens. The optimizer-assisted campaign reduced prompt redundancy and recursive recovery effort while preserving successful onboarding across all tested models. Overall, the results show that agentic AI can support zero-touch onboarding while exposing measurable differences among models in reliability, efficiency, and runtime convergence.