Ethan Mollick, a professor at the Wharton School of the University of Pennsylvania, said small performance gaps between generative AI models can become much larger once the work shifts from simple tests to tasks that require dozens of consecutive steps. Citing research on long-horizon execution, he argued that a modest gain in single-step accuracy can translate into a stronger ability to finish longer tasks, and that the planning and coordination role in an AI agent workflow should be handled by a model with better judgment.
The paper looks at execution, not idea generation
The paper, titled The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs, studies execution rather than whether a model can come up with a solution on its own. Researchers first supplied the knowledge and plan needed to complete a task, then observed whether the model could carry out the sequence reliably from start to finish.
According to the paper, even when smaller models get close to near-perfect accuracy on a single step, larger models can still execute more rounds correctly. As the number of steps increases, the step-by-step accuracy of all models declines. Small differences at the single-step level can keep accumulating over a long task.
Earlier mistakes can make later errors more likely
The gap is not explained only by longer context windows. The study said it observed a "self-conditioning" effect: when the conversation context includes a model’s earlier mistakes, the model becomes more likely to make additional errors later in the same run. Simply scaling up model size does not fully eliminate that effect.
The paper also said that giving a model time to think can reduce the problem and help it complete longer tasks within a single working session. The study has been published at ICLR 2026, and its latest version was revised in March this year.
Mollick extends the findings to agent workflow design
Mollick applied the research to model selection in agent systems. In his view, if a task involves research, tool use, result checking, and repeated revisions, then tiny differences in single-step capability can build up across the full workflow. On that basis, he argued that a stronger model should serve as the coordinator instead of assuming a cheaper model can replace a frontier model just because their scores look similar on short tests.
The article also makes clear that this is Mollick’s interpretation of how the findings may be used. The paper tested execution ability on specified long-horizon tasks. It did not directly compare multi-model agent architectures, and it did not quantify the cost-effectiveness of different model combinations. Companies deciding whether to use a more expensive coordinator model still need to test the question against their own tasks, error costs, latency, and inference expense.
Checkpointing and recovery still matter in deployment
The core challenge in long tasks is that errors can propagate through the chain instead of ending with a single wrong answer. In practice, breaking work into verifiable smaller steps, keeping human review points, and allowing the system to recover to an earlier state remain important ways to reduce risk.
The paper suggests that stronger models and extra thinking can extend the range of reliable execution, but that does not mean mistakes disappear. The takeaway is not that every job should run on the most expensive model. It is that benchmark results and deployment decisions need to account for task length and sustained execution ability. Similar scores on short questions do not automatically mean models are interchangeable inside complex workflows. Whether a stronger model should act as the coordinator still depends on whether a real workflow benefits from its judgment and oversight.