How to linearize a Simulink system with RL Agent

I have trained a reinforcement learning controller and it works fine. I want to linearize the system with and without the agent. However, I see the linearization gives unexpected results. The tim-domain simulation shows it is stable, but the linearized model results in an unstable system (many positive real-parts of the eigenvalues). Note: I see the agent's output is discrete (I know it from the plot which has a stair form) while the other part of the system is continous.

Any idea?

Thanks