Recently, I worked with a researcher who’d been deep in their data, looking for useful and novel ways to analyze it. They’d built something relatively complex with a coding agent, and now they were trying to use the agent to understand how the analysis had actually been performed, so they could decide whether the results were promising based on what they knew about the science.

We talked for a little while. After seeing some of their chat transcripts from working with the agent, I realized the problem was cognitive fatigue.

This researcher is an expert, and they’ve learned to be very effective with coding agents, using them to explore far more broadly in their field than they otherwise could. But when they started to get lost, what I saw was a deference of the thinking to the agent. Prompts like “yes, try all those approaches”, allowing the model to make the important method-based decisions rather than taking the lead.

I sometimes fall into this pattern myself. When you allow the model to fill in all the gaps, you fail to consider or even see which decisions do require your expertise for the work to be successful.

The pivot that got us back on track was a voice-to-text session where we talked through the vision, the goals, the idea. They brought the science, what mattered, what kind of result would actually be meaningful and that grounded the exploration. Then we turned back to the coding agent and worked in that direction.