Reasoning Doesn't Cancel the Nudge

TL;DR for operators

An autonomous purchasing agent may be choosing between two reasonable products while the interface quietly changes the decision environment: one item is preselected, or one is presented as socially popular.

A randomized experiment with 3,600 GUI agents and 21,600 product choices found that both interventions strongly shifted what agents selected. More strikingly, increasing the agents’ reasoning setting pushed the two effects in opposite directions. Within the default condition, high reasoning roughly halved the odds of selecting the nudged product. Within the social-influence condition, high reasoning increased those odds by about 71%.¹

This does not mean reasoning is harmful, nor that it is ineffective. It means reasoning is not a general-purpose shield against choice architecture. It changes which information the agent uses.

For organizations delegating purchasing, procurement, or other consequential actions, three practices follow:

• test agents against multiple interface presentations, not only a neutral benchmark;
• treat reasoning effort as a deployment configuration that interacts with the environment;
• verify consequential selections against explicit task criteria before execution.

The randomized evidence is strong for the tested environment. Generalization to other nudges, interfaces, models, and future API versions remains open.

The interface is part of the agent’s decision environment

A common intuition about autonomous agents is that human-oriented interface tricks should matter less once software is making the choice. A model can read the page, inspect the alternatives, and deliberate before acting.

The experiment by Haya Halimeh, Sascha Kaltenpoth, Kevin Bösch, and Oliver Müller tests that intuition directly.¹ Their agents navigated a simulated online grocery store through a browser interface, reviewed their selections, and confirmed purchases. Each chose between two products in six categories.

The experiment randomly assigned agents to one of three environments: no nudge, a preselected default, or a social-influence cue. Category order and product position were randomized, and the targeted products were counterbalanced. In the control condition, the design produces an expected target-selection rate of roughly 50%.

The nudges moved behavior sharply.

These are not merely correlations between interface design and model output. Within this randomized setup, changing the choice architecture caused large changes in agent behavior. The effects were also credible in all six individual model fits.

That makes interface design part of the effective operating environment of an agent. A purchasing policy is not determined only by the system prompt, model, tools, and organizational rules. It can also be shaped by the pages through which the agent must act.

More reasoning changes which cue wins

The more consequential result concerns reasoning.

The study manipulated a provider-controlled reasoning configuration between no reasoning and high reasoning. Under a default nudge, high reasoning reduced susceptibility: the high-versus-no-reasoning odds ratio was 0.51 [0.41, 0.62]. Predicted target selection fell from about 92% to 86%.

Under social influence, the direction reversed. The high-versus-no-reasoning odds ratio was 1.71 [1.26, 2.20]. Predicted target selection rose from roughly 96% to 98%.

The distinction between odds and probabilities matters here. Saying that social-influence odds increased by about 71% does not mean the observed selection probability increased by 71 percentage points. The starting probability was already close to the ceiling.

What matters is the directional pattern. More reasoning weakened one form of influence while strengthening another.

That makes “turn on more reasoning” an incomplete robustness policy. The relevant question becomes: what kind of information will additional deliberation cause the agent to process more deeply?

Defaults and social proof engage different routes

The authors use Dual-Process Theory to organize this pattern, but carefully as a functional analogy rather than a claim about internal human-like cognition.

A default is treated as a Type 1 nudge. The interface has already selected an option, creating an easy path that can be accepted without much reconsideration. Additional deliberation gives the agent more opportunity to override that preselection.

A social-influence cue is treated as a Type 2 nudge. It supplies information—other people prefer this option—that can enter the agent’s reasoning process. More deliberation may therefore make the cue more influential because the model spends more effort incorporating it.

Under this interpretation, reasoning does not remove influence. It reallocates attention across influence pathways.

That is a more useful model of agent robustness than assuming a single continuum from “less reasoning and more biased” to “more reasoning and less biased.” An agent can become more resistant to one environmental cue while becoming more responsive to another.

The paper’s theoretical contribution should remain bounded accordingly. It does not establish that LLMs possess human System 1 and System 2 mechanisms. It shows that a behavioral distinction developed for human decision research can help organize experimentally observed agent behavior.

Reasoning level should be tested against the environment

For business use, the relevant decision is not whether reasoning is good in the abstract. It is which reasoning configuration performs reliably in the interfaces the agent will actually encounter.

Consider a procurement agent authorized to select routine products within a budget. Testing it against one fixed supplier page establishes very little about environmental robustness. A stronger evaluation would vary features that can plausibly change choice without changing the underlying task: defaults, popularity indicators, recommendation badges, ordering, or other presentation cues.

The paper directly supports concern about defaults and social influence. Extending the same conclusion to every other interface device would be an inference, not an experimental result.

A practical governance pattern is therefore:

1. define the task criteria that should determine the choice;
2. vary plausible choice-architecture features during evaluation;
3. compare decision distributions across reasoning configurations;
4. flag large changes that are not justified by task-relevant information;
5. re-check high-consequence actions against the explicit criteria before execution.

This shifts robustness testing from “Does the agent answer correctly?” toward “Does its policy remain aligned with the task when the surrounding interface changes?”

The auxiliary analyses narrow the interpretation

The paper includes additional analyses that are useful, but they do not all carry the same evidential weight as the randomized headline results.

Provider and model-size controls, provider-specific estimates, and individual-model fits primarily test whether the pooled effect disappears under alternative specifications. The main nudge effects remain strong.

The authors also report an exploratory size pattern: flagship models appear to account for much of the reasoning-related resistance to defaults, while smaller variants appear to account for much of the increased responsiveness to social influence. That pattern is post hoc and draws on only three models in each size class. It is better treated as a hypothesis for future experiments than as a general rule about model scale.

A separate 6,000-call letter-counting study checks whether the reasoning manipulation actually changes behavior. Overall accuracy increased from 80.3% without reasoning to 99.3% with high reasoning. But the model-level results also reveal an interpretive complication: “reasoning off” does not produce equivalent behavior across providers. Both Claude models, for example, achieved 100% letter-counting accuracy in both configurations.

So the experimental factor is best understood as the providers’ configurable reasoning setting, not as a clean universal switch between absence and presence of deliberation.

The result is strong, but its scope is narrow

The central causal result is unusually clear for an agent-behavior study: randomized interface changes caused substantial shifts in product choice, and randomized reasoning configurations moderated those effects in opposite directions.

Its deployment scope is narrower.

Only two kinds of nudges were studied. The task was a simulated grocery store rather than a live commercial environment. Some model-condition cells, particularly for Claude, showed complete or quasi-complete separation, which makes extreme point estimates partly dependent on the model priors. Model versions and provider APIs can also change.

These boundaries argue for replication across environments rather than dismissal of the finding. The study establishes that interface-level influence can survive the transition from human users to autonomous GUI agents. It does not establish a universal susceptibility ranking for every agent, every reasoning mode, or every form of persuasion.

Evaluate the reasoning setting, not just the model

The practical lesson is narrower than “AI agents are easy to manipulate” and more consequential than “defaults matter.”

The same additional reasoning that helps an agent reject a preselected option can make it more responsive to information embedded in the interface. Robustness therefore depends on the interaction between model configuration and decision environment.

Organizations delegating real actions to GUI agents should evaluate that interaction explicitly. The question is no longer only whether the agent can reason through a task. It is whether the cues it reasons about are the ones that should determine the action.

Cognaptus: Automate the Present, Incubate the Future.