Sixth-generation (6G) edge networks require adaptive wireless environments that can react to channel variation, interference, and localized device constraints. Intelligent reflecting surfaces (IRSs) provide a programmable propagation layer, but conventional IRS controllers are often static, optimization-heavy, or weakly coupled to semantic network intent. This article presents LAMI, a closed-loop agentic metasurface intelligence framework in which edge-side small language models (SLMs) summarize local link conditions and an IRS-side large language models (LLM) agent generates bounded control directives for metasurface reconfiguration. Unlike open-loop LLM-assisted IRS control, LAMI couples semantic reasoning with deterministic validation, confidence gating, and fallback control before physical actuation. Under the evaluated controller-side simulation setup, LAMI reduces energy per decision by 25.7% and controller- side latency by 20.3% compared to a standalone LLM-IRS baseline, while improving SNR recovery and distance-adaptive response. We also discuss latency accounting, trustworthiness, scalability limits, and reproducibility artifacts for practical evaluation.