The deployment of compact large language models on smartphones enables on‑device autonomous decisions, such as emerging mobile GUI agents. Yet many decisions, like network configuration when apps switch or connectivity changes, must meet strict latency constraints to preserve user experience.\ \ We introduce PEAM, an on‑device LLM‑based communication agent that autonomously generates personalized network‑configuration rules executable within milliseconds. PEAM first fine‑tunes the on‑device LLM with a dataset built from manually annotated real‑user data, embedding communication knowledge into the local model.\ \ PEAM operates in two phases. In the rule‑generation phase, run during idle periods (e.g., overnight), it batch‑processes historical network‑configuration records, aligns evidence across batches, reduces local bias, and derives robust configuration rules. To improve accuracy over time, PEAM constructs a user profile from high‑confidence rules, capturing the user’s network preferences and feeding this profile into each generation cycle.\ \ In the rule‑execution phase, triggered by changes in communication context (such as app switches or connectivity shifts), PEAM searches the rule library, selects the appropriate rule, and executes it within milliseconds.\ \ Experiments show that PEAM, based on a fine‑tuned 4B model, achieves 99.13% correct network‑configuration decisions, with 92.50% of generated rules being valid. Compared with a traditional expert‑rule method, it improves correctness by 4.41% and validity by 84.50%.\ \ Review