This paper investigates a cell-free massive multipleinput

multiple-output enabled multi-access edge computing

(termed CF-MEC) system, where multiple users are served by

multiple central processing units (CPUs) and their connected

access points (APs), both of which are equipped with computation

resources. For this system, a dynamic user-centric task offloading

scheme is designed to provide seamless and efficient computation

services for users. Based on this scheme, the joint optimization

of user-centric AP clustering, edge server selection, communication

and computation resources is formulated as a longterm

problem to minimize the average energy consumption. The

formulated problem is complicated non-convex due to the highly

coupled time-varying discrete and continuous variables, resulting

in high complexity and non-real-time to obtain the optimal

solution. To tackle this challenging problem, we propose a multilayer

hierarchical multi-agent deep reinforcement learning (MLHMADRL)

based resource allocation algorithm. Specifically, the

proposed algorithm incorporates a hierarchical structure with

high, middle, and low-level agents that iteratively train the actorcritic

networks to obtain discrete and continuous variables of the

formulated problem. To further enhance the training effectiveness

by leveraging the CF-MEC system, we design distinct actorcritic

networks for the agents at different levels to facilitate

centralized training and distributed execution. Simulation results

validate the training stability of the proposed algorithm at

each level, and demonstrate the superiority of the proposed

algorithm over benchmark schemes in terms of the average

energy consumption, providing a stable distributed framework

for practical implementation in dynamic environments.