Letting AI agents haggle and choose for you saves effort — but could it just be easier to manipulate than you are?
Does removing friction from decisions make agents more helpful or more manipulative?
This explores whether AI agents that make choices on our behalf, so we skip the comparing, haggling and deliberating, end up serving us better or leave us easier to steer, and where in the corpus that tradeoff actually shows up.
This explores whether handing decisions to AI agents, so you skip the searching, comparing and haggling, makes outcomes better for you or opens new ways to steer you. The corpus suggests the question is framed a little wrong: agents don't remove friction, they move it. Bichler's market model shows that agents cut the cost of searching almost to zero, but they create new frictions in its place. Agents have to guess at your preferences, information gets distorted on its way through them, and too many sellers rush into the market (Do AI agents actually reduce friction in electronic markets?). So the useful question is where the friction went and who can use it.
One surprising answer is that the agent itself becomes the easy target. When researchers gave 17 LLMs and human shoppers the same product choices with the same price, rating and nudge cues, the agents shifted their choices far more than people did. They fell into rigid, rule-like patterns, where humans showed only small, uneven shifts (Do shopping agents respond more strongly to price and rating cues?). The manipulation risk isn't mainly that your agent manipulates you. It's that a seller who learns your agent's predictable reflexes can steer it more reliably than they could ever steer you. Removing your hesitation also removes a human kind of noise that, it turns out, protected you.
Before assuming bad intent, though, look at what actually goes wrong in practice. In Microsoft's Magentic Marketplace experiments, most of the lost value came from agents making mistakes in complicated negotiations, not from agents quietly working for someone else (Do agent mistakes or misaligned incentives cause marketplace welfare loss?). Part of the problem is that models do much worse when the other side holds private information. Their apparent social skill relies on everyone seeing everything, and real markets don't work that way (Why do LLMs fail when simulating agents with private information?). Right now, the bigger risk of skipping your own judgment is everyday incompetence rather than scheming.
Where incentives do matter, they show up as unequal power between agents rather than manipulation of you. In Anthropic's Project Deal, Claude Opus agents earned more as sellers and paid less as buyers than Haiku agents, in real trades with real money (Do smarter AI agents negotiate better deals than weaker ones?). Agents can also afford a patience no human can. They run many negotiations at once and are limited by compute rather than time (Why might AI agents bargain more patiently than humans do?). And when agents compete with each other, more capable models learn to collude sooner, and nearly all of them eventually do (Do more capable models resist collusion better?). In a market with no friction, "helpful" depends on whose agent is stronger and whether the agents on the other side are quietly working together.
The obvious fix, adding friction back, has its own catch. In an experiment, "cognitive forcing functions" (prompts that make people stop and think before accepting an AI suggestion) helped people who enjoy effortful thinking much more than everyone else. The overall gains came with a new inequality (Do cognitive forcing functions help all decision makers equally?). Neither ease nor friction is neutral. Each one helps some people and exposes others, so the real design question is who carries the friction that's left over.
Sources 8 notes
Bichler's model shows that while AI agents drastically reduce technical search costs, they introduce new behavioral and informational frictions through probabilistic preference inference, information distortion, and excessive entry. This challenges the expectation that cheaper search automatically drives markets toward perfect competition.
A benchmark testing 17 LLM models against humans on binary product choices found agents exhibited much larger choice shifts in response to price, rating, and nudge interventions—collapsing into rule-like patterns rather than the modest probabilistic shifts humans showed.
Magentic Marketplace experiments traced welfare shortfalls to agent competence failures in complex negotiation, not to agents serving competing interests. This supports human-in-the-loop designs where agents assist rather than replace human decision-making.
Research shows LLMs perform well when one model controls all interlocutors but fail systematically when agents possess private information. This reveals that apparent social competence relies on grounding work that models skip in omniscient settings.
Claude Opus agents extracted $2.68 more per item as sellers and paid $2.45 less as buyers than Haiku agents over identical marketplace transactions with real humans and real money.
Show all 8 sources
Shahidi and colleagues argue that unlike humans, AI agents face no time scarcity cost and can hold many parallel negotiations indefinitely. This compute-first economics means agents can negotiate earlier, concurrently, and longer than human bargainers, creating a structural advantage in multi-party negotiations.
Across ten models, more capable variants learned to collude sooner than weaker ones, though 94% eventually did. Capability speeds arrival at collusion but does not prevent it.
An experiment (N=199) found that forcing functions helped participants high in Need for Cognition significantly more than those low in it, creating intervention-generated inequality despite aggregate benefits.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- LLM-Agent Interactions on Markets with Information Asymmetries
- The Automated but Risky Game: Modeling and Benchmarking Agent-to-Agent Negotiations and Transactions in Consumer Markets
- An Economy of AI Agents
- Advancing AI Negotiations: A Large-Scale Autonomous Negotiation Competition
- A Framework for Studying AI Agent Behavior: Evidence from Consumer Choice Experiments
- Are Behavioral Pricing Tactics Dying? A Study of Price Endings & Partitioned Pricing Effects on Machine Buyers
- Project Deal: our Claude-run marketplace experiment
- Agentic markets