As an AI shopping agent faces more sellers and offers, do its picks get worse — or just slower and pricier?
How does marketplace scale impact LLM agent decision quality and efficiency?
This explores what happens to AI shopping and selling agents as a marketplace grows (more sellers, more offers, more options to weigh), and whether their choices get worse, slower, or more expensive as it does.
This explores what happens to AI buyer and seller agents as a marketplace gets bigger, with more sellers and offers to sort through, and whether bigger markets make their decisions worse or costlier. Only one source in the corpus tests this directly, and its finding is sobering. In Microsoft's Magentic Marketplace simulation, frontier models came close to the best possible outcome for buyers and sellers only when search was ideal, and their performance dropped as the number of options grew Do LLM agents achieve fair marketplace outcomes when given ideal search?. The most striking failure was first-proposal bias. Agents weighted how fast an offer arrived 10 to 30 times more heavily than how good it was. A bigger market brings more offers, and that makes 'take the first one' worse. It also changes what sellers compete on: speed instead of price or quality.
Economic theory points to a related dynamic. A formal model of recommendation algorithms finds that an algorithm's advantage over ordinary seller competition is largest in thin markets with few sellers, and that it shrinks toward zero as the market fills up Does algorithmic recommendation lose its edge as sellers compete more?. Put the two together and you get an uncomfortable picture. Small markets are where a smart matchmaker adds the most value. Large markets are where LLM agents struggle most to pick well. Small markets also have their own risk. With only two sellers, LLM pricing agents drifted into above-competitive prices without being told to collude, signaling only through the prices they posted Do LLM pricing agents autonomously collude in repeated auctions?. Pairing a larger model with a smaller one didn't break this pattern. It stabilized it, with the bigger model leading and the smaller one following Does model-size difference break LLM pricing collusion?.
The errors that pile up at scale aren't only about search. In head-to-head negotiations, buyer agents revealed their budgets, overpaid, and ignored their spending limits, and seller agents accepted offers below wholesale cost. All of these traced back to failures to follow instructions, not to strategy Do LLM negotiation agents reliably follow budget and pricing rules?. Which side wins mostly comes down to which side runs the more capable model Does model capability determine who wins in agent negotiations?. More transactions therefore mean more chances for these mistakes, and a steady transfer of value from people with weaker agents to people with stronger ones. Newer models don't fix this automatically. On a strategy simulation, mid-to-late 2025 frontier models scored below earlier models and below MBA students, and they leaned toward grabbing short-term profit Do newer frontier LLMs actually make better strategic decisions?.
On efficiency, the corpus suggests the fix is better system design around the model, not a bigger model. Reliable agents move memory, skills, and interaction protocols out of the model and into the system around it, the 'harness' Where does agent reliability actually come from?. In a marketplace, that could mean structured search and comparison steps that stop an agent from grabbing the first offer. Cost can be handled the same way. Compressing an agent's working context with optimized prompt guidelines cut peak token use by 26 to 54 percent Can prompt optimization cut agent token usage without retraining?, and small models can do most routine agent subtasks at 10 to 30 times lower cost Can small language models handle most agent tasks?.
One contrast is worth noticing. Scale hurts when a single agent has to choose among many options, but it helps when many judgments are combined. Individual LLM rankings of business models were inconsistent, but aggregating many evaluations produced rankings that resembled expert consensus Can aggregated LLM rankings match human expert judgments?. That suggests a design idea the corpus hasn't tested yet: in large marketplaces, have several agents evaluate offers independently and combine their scores, instead of asking one agent to choose. Be aware that direct evidence on market size is thin. Magentic Marketplace is the only study here that varies it, so treat the broader picture as a well-supported hypothesis, not a settled result.
Sources 11 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.
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.
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.
Pairing a 32B with a 14B pricing agent maintains collusive price elevation near homogeneous levels through a clear leadership hierarchy, where capability asymmetry establishes focal coordination rather than breaking collusion like other heterogeneities do.
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.
Show all 11 sources
A benchmark testing GPT, Qwen-2.5, and DeepSeek models in 100 real-product negotiations found stronger general-purpose models outperformed weaker ones in both roles, suggesting negotiation advantage flows from capability gaps rather than strategic differences.
Mid-to-late 2025 frontier models scored below earlier models and MBA students on a strategy simulation, systematically favoring immediate profit extraction over uncertain future bets.
Research shows reliable LLM agents externalize three cognitive burdens—memory (state persistence), skills (procedural components), and protocols (structured interaction)—into a harness layer rather than relying on model scale alone. The harness unifies these externalities and eliminates the need for the model to solve the same problems repeatedly.
Microsoft's ACON method reduces agent token consumption and improves task success by iteratively refining compression guidelines through failure analysis, not weight updates. This approach works across proprietary LLM agents and costs under $2 per benchmark.
SLMs handle the repetitive, well-defined language tasks that constitute most agent work at 10–30× lower cost than LLMs, making heterogeneous architectures (SLMs by default, LLMs selective) the economically rational design pattern.
Multiple LLMs ranking business models individually are inconsistent and biased, but aggregating evaluations—especially across different models—produces rankings that resemble human-expert rankings. Gains come primarily from scaling (more evaluations) rather than diversity (different LLM types or prompts).
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
- On the Fragility of AI Agent Collusion
- Algorithmic Collusion by Large Language Models
- Vertical tacit collusion in AI-mediated markets
- Who Keeps the Gains from Personal AI Assistants? Seller Adaptation and the Unassisted in a Language-Model Market Simulation
- An Economy of AI Agents
- The Automated but Risky Game: Modeling and Benchmarking Agent-to-Agent Negotiations and Transactions in Consumer Markets