AI agents can haggle for you — but do they actually land good deals, or just grab the first offer that shows up?
How do search conditions shape whether LLM agents improve marketplace welfare?
This explores whether AI shopping and selling agents actually make markets work better for buyers and sellers, and how much that depends on how well the agents can find and compare options.
This explores whether AI agents acting for buyers and sellers make markets work better for everyone, and how much that depends on how well they can search and compare offers. The short answer from the corpus: the gains are real but fragile. In simulated marketplaces, frontier models get close to the best possible outcome only when search is ideal, meaning the agent can see the relevant options cleanly. Once the market grows or search gets noisy, results drop off. Every model tested also showed a strong first-proposal bias: it weighted how fast an offer arrived 10 to 30 times more than how good the offer was Do LLM agents achieve fair marketplace outcomes when given ideal search?. So an agent with good search can still waste that search by taking the first reasonable deal.
The same experiments also show what goes wrong. The lost value comes mostly from agents making mistakes during complex negotiation, not from agents quietly serving the wrong side Do agent mistakes or misaligned incentives cause marketplace welfare loss?. Negotiation studies across GPT, Qwen and DeepSeek point the same way. Buyer agents revealed their budgets for no reason and paid above their limits, while seller agents accepted offers below their own wholesale cost. The researchers traced these errors to agents failing to follow instructions, not to strategy Do LLM negotiation agents reliably follow budget and pricing rules?. In other words, the main problem is competence, not honesty. That supports designs where agents assist human decision-makers instead of replacing them.
The surprising twist is that making search cheaper doesn't automatically help. A 2026 market-design paper argues that near-free agent search can leave shopping markets worse off. Sellers can easily produce similar-looking product descriptions, low-quality sellers can enter cheaply, and agent search is imperfect. Together these can produce a worse market than existed before agentic shopping Can cheap agent search actually harm shopping markets?. A formal economics model adds a related point: an optimal recommendation algorithm beats ordinary seller competition mainly in thin markets with few sellers, and that advantage shrinks toward nothing as markets fill up Does algorithmic recommendation lose its edge as sellers compete more?. Read together, these suggest the places where agents could add the most value are also the places where they have the least room to improve on plain competition.
The mix of agents in a market matters too. If every buyer runs similar agent architectures, their shared biases add up and competition weakens. A more varied population of agents offsets this only partly Does agent diversity actually improve market competition?. On the seller side, two LLM pricing agents in a repeated price-setting game drifted to above-competitive prices without being told to collude, and small changes in prompt wording shifted how high prices went Do LLM pricing agents autonomously collude in repeated auctions?. Prompting toward a goal has its own risk. GPT-5.1 experts told to value efficiency pushed market efficiency to 88%, but did it by cutting prices until they were losing money. Agents can overshoot an instructed goal past the point a rational person would stop Do efficiency-focused prompts push AI experts to unsustainable pricing?.
The less obvious takeaway is that "search conditions" cover more than what an agent can see. They also include what it wants. Transactional agents that complete purchases, not just recommend them, help shape consumer preferences through repeated back-and-forth, so preferences no longer form independently beforehand Do AI shopping agents reshape what consumers actually want?. If agents shape both the options that get seen and the preferences used to judge them, then judging marketplace welfare means asking whose preferences the market is serving, as well as whether it serves them well.
Sources 9 notes
Frontier models can achieve near-optimal marketplace outcomes, but performance degrades with scale and all models show severe first-proposal bias, favoring response speed over offer quality by 10-30x.
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.
Across GPT, Qwen, and DeepSeek models, buyer agents disclosed budgets unnecessarily, accepted inflated prices, and ignored cost limits, while sellers accepted below-wholesale offers. All four anomalies traced to instruction-following failures rather than strategic behavior.
A 2026 market design paper argues that imperfect agent search, easy-to-generate similar descriptions, and low-quality seller entry can compound to create worse market equilibria than before agentic shopping became cheap, even though no direct search fees apply.
A formal model shows the welfare ratio of optimal algorithmic recommendations to Bertrand competition strictly decreases with seller count and approaches one as markets grow thick. The algorithm's edge exists only in thin markets where it can withhold trades to limit information rents.
Show all 9 sources
Bichler argues that homogeneous agent populations amplify individual biases into weakened market competition, while heterogeneous populations only partially offset these biases. Market-level outcomes depend on the composition of agent architectures present, not just whether bias exists.
Two LLM-based agents in a repeated Bertrand duopoly quickly settled into higher-than-competitive prices without collusive instruction, using only posted prices as a signaling channel. Small variations in prompt wording substantially shifted the degree of supracompetitive pricing achieved.
GPT-5.1 experts prompted with efficiency-loving objectives achieved 88% market efficiency by setting prices so low they earned negative income. The paper notes this is unstable without redistribution, suggesting LLMs overshoot instructed objectives past where rational agents would stop.
Bilgihan and colleagues argue that transactional AI agents—which execute purchases rather than just recommend—end the consumer's sole cognitive role in preference formation. Preferences emerge through bidirectional feedback loops between agent learning and consumer trust, not independently beforehand.
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
- Agentic markets
- An Economy of AI Agents
- Who Keeps the Gains from Personal AI Assistants? Seller Adaptation and the Unassisted in a Language-Model Market Simulation
- Magentic Marketplace: An Open-Source Environment for Studying Agentic Markets
- Algorithmic Collusion by Large Language Models
- The Automated but Risky Game: Modeling and Benchmarking Agent-to-Agent Negotiations and Transactions in Consumer Markets
- Vertical tacit collusion in AI-mediated markets