This paper addresses the Multi-Agent Pickup and Delivery (MAPD) problem, a dynamic extension of the Multi-Agent Path Finding (MAPF) problem, critical for applications like automated warehouses where agents perform continuous pickup and delivery tasks. Unlike traditional MAPD solutions assuming idealized conditions, our approach tackles real-world challenges like agent delays, uncertainties, and dynamic task allocation. Existing methods often exhibit inefficiencies in task assignment and idle agent management, limiting practical applicability. We propose a robust framework integrating token-based task allocation, adaptive path planning, and a global proximity-based task assignment strategy to optimize efficiency. The framework handles disruptions caused by unknown delays or deadlocks and reallocates non-task endpoints to manage idle agents effectively. Extensive simulations validate the framework, demonstrating its superiority over state-of-the-art algorithms like TP, TPTS, and k-TP in maintaining efficient, collision-free operations and adapting to runtime uncertainties. This scalable solution advances real-world multi-agent coordination in domains such as warehouse automation and autonomous fleets.