I buy into trading raw concurrency for wall-clock time. Having thousands of agents run for 32 minutes to replace months or a full year of human work makes total sense. Massive piles of experimental data sit idle because no researcher can comb through decades of academic papers. The workflow works because the task is sliced fine enough, with each agent assigned to dig into just one genetic perturbation.
In blind evaluations, domain experts preferred the agent-generated reports 75% of the time. When you lock down the boundaries, distributed inference can wake up mountains of archived data that would otherwise collect dust. Spending the compute tokens here is completely worth the cost.