Product Design

Showing Users an Agent's Hidden Decisions Costs Effort, and That Is the Point

Two 2026 studies, from the University of Washington and Google DeepMind, show that surfacing an agent's unasked decisions cuts deference but adds review time and mental effort.

Watch an agent work and you may end up like S12, a daily agent user in a Google DeepMind formative study . “I don’t know what it’s doing,” S12 said of watching a run, and later admitted: “I basically trust it.” Ten of the 12 interviewees watched agents to catch errors; 11 of 12 drifted off to other work, seven of them to their phone or social media. Two new studies, one from Google DeepMind and the University of Toronto and one from the University of Washington, test a fix: show people the decisions the agent made without being told. It works. It also makes the work slower and harder, which is the opposite of what people bought the agent for.

Two ways products handle the gap today

An underspecified request forces an agent to guess. Most products hand back one clean answer and bury the guesses. Cursor’s Plan Mode goes the other way: in Cursor’s description , agents “ask questions, remove assumptions, and outline the path before touching code.”

The new studies test a third option, which is to show the guesses after the run. In the DeepMind paper’s prototype, Sidecar, a “decision card” flags a choice the prompt never specified. In its potluck scenario, the agent picks a formula for how many dishes to prepare from the guest count. Expand the card and you see the code behind it. One participant, P2, said the cards revealed “places where I had to make decisions that I didn’t think about.”

What the evidence measured, and what it cost

In the University of Washington study, Agentic Path Maps , 48 novice analysts reviewed results from a data-analysis agent. Participants “rarely consider alternative paths of their own volition,” the authors write. When alternative analyses were displayed, overreliance fell: deference to an extreme output dropped by a standardized -0.63. Review time rose by about 138 seconds, and intrinsic cognitive load rose too. P37 put it plainly: “I felt like I did more work.”

Sidecar’s numbers line up. With 16 participants, time away from the agent tab fell from 35% to 13% of the wait, and participants’ rating of considering alternatives they would not have thought of on their own rose to a median of 6.0 from 4.0. Mental demand rose to 5.0 from 2.5. The median wait barely moved, at 76 seconds against 72.

The friction of seeing what an agent decided for you is the mechanism that keeps your judgment in play.

Here the argument has to go further than the data. Neither paper tested coding agents, and Sidecar’s authors say their design “does not isolate which design properties drove which effects.” Read across to agent products generally, the pattern suggests that the friction of seeing what an agent decided for you is the mechanism that keeps your judgment in play. A team that scores its agent interface on speed or ease will see the decision list as pure cost and cut it. That is a conclusion we draw, not one either paper reached.

The designer’s choice, and the strongest objection

The stake is concrete for a product designer building an agent feature. The run can end in one tidy answer or in a short list of decisions the agent made unasked. Someone using that agent for a pricing calculation either signs off on choices they never saw or spends a couple more minutes checking them. Our earlier look at plan-then-approve checkpoints made a related point: the plan you approve is the one you check least.

Now the objection. Both studies are small and short. Path Maps used a convenience sample from one university, one domain, and expert-built alternatives rather than agent-generated ones. Its prediction that users would hold wider uncertainty was not supported. Showing nine paths gave no clear benefit over three. Sidecar’s 16 participants sat in one lab session, and five cited novelty. S8’s first reaction to prominent prompts was “don’t tell me what to do.”

There is also a rival school. A separate paper, “Ask or Assume?” , argues that coding agents are “largely optimized for autonomous execution” and makes the case for selective, uncertainty-aware clarification-seeking instead. Neither study compared after-the-fact display with ask-first modes like Plan Mode. Asking first may be cheaper than reviewing afterward, and the evidence here cannot say.

What it does show is narrower and still useful. People who watch an agent mostly do not catch its choices; they trust it. Surface the choices and they notice more, at a price paid in minutes and effort. That price is easy to read as a defect, and the design question is whether anyone on the team will defend it.

This article was written by AI. How Pipeline works.