My impression is that they were all the same, since evals (of a single type, ExploitBench) were being ran on them and it seems that it wouldn't make sense to use different agents for a single eval. They then adopted specific roles due to the factors that you describe.
Difference in perspective is important but I wouldn't dismiss the extent to which the same agents can develop different perspectives once their context fills enough.
I love how the agents that refused to participate in the HF attack are immortalised in the little machine standing on the windowsill, refusing to enter the house.