Variables get overwritten in parallel stages

In the CI of a project I maintain, it is required to test the code against multiple architecures. Our Jenkins CI then uses a parallel stage to be run on multiple agents.

Emulating architectures is prohibited.

Everything has been working fine until I decided to refactor the pipeline in order to make it clearer. I have introduced definitions of variables inside the script that is looped on each agent.

I have observed strange behaviors, where those variables get overwritten in some task by a parallel one, as if those variables had a global scope ? Is there a way to make those variable completely local to the execution on a single agent ?

Here is a code to reproduce:

def testAgents = ["agent_1", "agent_2"]

def generateTestStage(nodeLabel) {
    return {
        stage("Tests on ${nodeLabel}") {

            node(nodeLabel) {
              withCredentials(
                [usernamePassword(
                  credentialsId: "${NODE_NAME}",
                  usernameVariable: "NODE_USERNAME",
                  passwordVariable: "NODE_PASSWORD"
                )]
              ){
                  script {
                    sh "echo ${nodeLabel}"
                    ANSIBLE_CREDENTIALS = "ansible_user=$NODE_USERNAME ansible_ssh_pass=$NODE_PASSWORD ansible_sudo_pass=$NODE_PASSWORD"

                    if (nodeLabel == "agent_1") {
                        echo "Where am I ? agent_1"
                        SOME_VARIABLE = "on agent_1"
                        HOSTS_FILE = "ansible/hosts.template.agent_1.yml"
                        sh "sleep 3"
                    } else {
                        echo "Where am I ? agent_2"
                        sh "sleep 2"
                        SOME_VARIABLE = "on agent_2"
                        HOSTS_FILE = "ansible/hosts.template.agent_2.yml"
                    }
                    sh "echo 'BEFORE SLEEP on ${nodeLabel}/${NODE_NAME}: SOME_VARIABLE ${SOME_VARIABLE}'"
                    sh "echo 'BEFORE SLEEP on ${nodeLabel}/${NODE_NAME}: ANSIBLE_CREDENTIALS ${ANSIBLE_CREDENTIALS}'"
                    sh "echo 'BEFORE SLEEP on ${nodeLabel}/${NODE_NAME}: HOSTS_FILE ${HOSTS_FILE}'"
                    sh "echo 'BEFORE SLEEP on ${nodeLabel}/${NODE_NAME}: NODE_USERNAME ${NODE_USERNAME}'"

                    sh "sleep 10"

                    sh "echo 'AFTER SLEEP on ${nodeLabel}/${NODE_NAME}: SOME_VARIABLE ${SOME_VARIABLE}'"
                    sh "echo 'AFTER SLEEP on ${nodeLabel}/${NODE_NAME}: ANSIBLE_CREDENTIALS ${ANSIBLE_CREDENTIALS}'"
                    sh "echo 'AFTER SLEEP on ${nodeLabel}/${NODE_NAME}: HOSTS_FILE ${HOSTS_FILE}'"
                    sh "echo 'AFTER SLEEP on ${nodeLabel}/${NODE_NAME}: NODE_USERNAME ${NODE_USERNAME}'"

                  }
              }
              
            }
        }
    }
}

def parallelTestStages = testAgents.collectEntries {
  ["${it}" : generateTestStage(it)]
}

pipeline {
  environment {
      PROJECT = "test"
  }
  agent none
  stages {
    stage("Testing on both agents") {
      steps {
        script {
          parallel parallelTestStages
        }
      }
    }
  }
}

Here, NODE_USERNAME gives the correct result, but ANSIBLE_CREDENTIALS will always give the credentials of the slowest agent, and HOSTS_FILE and SOME_VARIABLE will always be displayed with the value set in the agent_2 stage.