Parallel Stepping Conflict

mesa-llm enables parallel stepping by patching Mesa's AgentSet.shuffle_do.
When model.agents.step() is executed, it creates a single flat batch of all agents, which are then run concurrently.

The Conflict

The scheduler treats a MetaAgent M and one of its constituent agents A as independent peers and schedules them simultaneously.

If M internally triggers its agents:

for agent in self.agents:
    agent.step()

then A will be stepped twice in the same model tick:

1. Once by the parallel scheduler
2. Once by its parent MetaAgent

This leads to inconsistent state updates.

Fix: Create a "Team Badge" (is_component)

Introduce a small internal flag on agents called is_component.

self.is_component: bool = False

Meaning:

• False → Independent agent (scheduled normally)
• True → Component of another agent (not scheduled directly)

Behavior

The scheduler ignores agents where:

agent.is_component == True

This ensures that component agents are only stepped by their parent MetaAgent, preventing duplicate execution in a single tick.