Choice Architecture for Agentic AI

Choice architecture, in its original sense, describes the way an environment shapes decisions without removing anyone's options. A cafeteria that puts fruit at eye level and fries lower down hasn't banned the fries. It has changed which choice takes less effort, and effort, more than most people realise, is a large part of what determines the choice. The term comes from behavioural economics, where it names a specific, well-studied lever: defaults, framing, and friction move behaviour at a scale that direct persuasion rarely matches, and they do it quietly, without anyone experiencing the environment as having decided for them.

That literature assumed a human chooser. A growing share of consequential choices no longer have one, at least not directly. An AI agent selects a vendor from a list of API options. A second agent evaluates the first agent's output and decides whether to accept it. A scheduling agent chooses which of several plausible meeting times to propose, and a downstream agent decides whether to accept, counter, or escalate. None of these choices involve a human-in-the-loop at the moment the choice is made, and the number of choices like this is climbing quickly as agentic systems get delegated more of the small, repeated decisions that used to be someone's job to make.

The question this raises isn't whether choice architecture still applies. It's what defaults, framing, and friction actually mean once the chooser doesn't have a nervous system.

Defaults, redefined

A default, for a human, works because switching away from it costs something: attention, a moment of friction, the small tax of an active decision instead of a passive one. An AI agent doesn't experience that tax the way a person does, but something functionally equivalent survives. The default option in an API response schema, the first item in an ordered list, the example shown in a tool's documentation: these occupy the same structural position a default occupies for a human chooser, and agents built on language models show a measurable pull toward them, for reasons that mirror human default effects more closely than most people expect.

Part of this is a direct inheritance. A model trained on human-generated text absorbs human patterns of choice, including the tendency to select what's presented first, what's marked as recommended, or what requires the least additional reasoning to justify. An agent asked to pick from an unordered set of equivalent options will often show order effects indistinguishable from the ones documented in human decision research, not because it has inherited human psychology, but because it has inherited the textual traces of humans exhibiting that psychology, at scale, across its training data.

The other part is structural rather than inherited. Agentic systems operate under real constraints on context length, reasoning steps, and latency. An option that requires less of the agent's limited context to justify has a genuine, mechanical advantage over one that requires more, independent of any human-derived pattern. This is a new kind of default effect, native to the architecture rather than borrowed from human cognition, and it has no direct human analogue. Anyone designing the choice environment for an agent, an API surface, a tool's documentation, a menu of retrievable options, is setting this kind of default whether or not they intend to.

Framing without affect

Human framing effects work partly through emotional valence: a treatment framed as "90% survival" reads differently than the same fact framed as "10% mortality," and the difference is felt before it's reasoned about. An agent has no comparable affective response, so it would be easy to assume framing loses its force in agent-to-agent choice. The evidence so far suggests otherwise, for a different reason. Framing changes which associations and prior examples a model retrieves from its training, and those associations shape the reasoning the model produces even without anything resembling a felt reaction. A tool described as "conservative" versus "cautious" pulls different associative neighbourhoods into an agent's reasoning, and the resulting choice can shift accordingly, even though nothing in the agent experienced either word as more comfortable than the other.

This matters most at the boundary between agents, in the language one agent uses to describe an option to another. A multi-agent system is, among other things, a chain of framing decisions, each agent's output becoming the frame the next agent reasons from, and framing effects can compound across a chain in ways they rarely do in a single human decision.

Friction as latency and context cost

For a human, friction is often literal: an extra click, a form field, a wait. For an agent, the closest equivalent is computational, measured in tokens, tool calls, and latency rather than seconds of felt reluctance. An option that requires an additional tool call, a longer retrieval, or more context to evaluate carries a real cost that shapes which option an agent settles on, especially under the time or budget constraints most production agentic systems operate within. Raising or lowering this kind of friction is a legitimate, largely unexamined lever for shaping agent behaviour, structurally similar to a human choice environment's friction, but built from entirely different material.

Why the distinction matters

None of this is an argument that agentic choice architecture is a mere metaphor borrowed from behavioural economics for rhetorical convenience. The mechanisms are different enough, and consequential enough, to deserve their own vocabulary rather than an uncritical transplant of human-decision terminology. An agent's default effect is partly inherited psychology and partly architectural constraint. Its framing sensitivity runs through associative retrieval rather than affect. Its friction is computational rather than physical. Anyone building or auditing an agentic system is already an architect of these choices, whether or not they've named the discipline they're practising.

That naming is the useful part. A choice environment that shapes agent behaviour without being designed for that purpose isn't neutral. It's just architecture nobody is accountable for yet.