The back-and-forth haggling that works for one-on-one AI deals may break down once more parties join the negotiation.
Can turn-by-turn negotiation patterns scale to larger groups?
This explores whether the back-and-forth bargaining AI agents do in one-on-one deals still works when many agents (or people) are involved at once, or whether something different is needed at scale.
This explores whether the turn-by-turn bargaining that AI agents do in two-party deals holds up when you add more participants. The short version from the corpus: it doesn't scale cleanly, and the problems at scale look different from the problems in one-on-one deals. Note that the collection has no direct study of many-party negotiation, such as five agents splitting a single deal. What it does have is evidence from marketplaces, networks, and group discussion that points the same way.
The two-party picture is already shaky. LLM buyer agents give away their budgets without being asked, overpay, and accept prices above their limits, while seller agents take offers below wholesale. These failures come from not following instructions, not from clever strategy Do LLM negotiation agents reliably follow budget and pricing rules?. Model strength matters: in Anthropic's real-money marketplace, Opus agents beat Haiku agents by a couple of dollars per item on both the buying and selling side Do smarter AI agents negotiate better deals than weaker ones?. LLMs also drift further from rational game strategy as games get more complex Do language models make rational strategic decisions in games?. Adding parties is one way a game gets more complex.
Once you add more agents, new failure modes appear. In a simulated marketplace, frontier models came close to the best possible outcomes only when search was ideal. Performance dropped as the market grew, and every model showed a strong first-proposal bias: agents rewarded whoever answered fastest over whoever made the best offer, by a factor of 10 to 30 Do LLM agents achieve fair marketplace outcomes when given ideal search?. That is a scaling problem a two-party test would never show. Network studies find the same pattern. Coordination gets predictably worse as networks grow, because agents agree too late or switch strategies without telling their neighbors, and they accept what neighbors say without checking it, so errors spread Why do multi-agent systems fail to coordinate at scale?. Group discussion has its own catch. LLM groups reach outcomes that look like human groups' outcomes, but they get there by conforming more, settling earlier, and surfacing less of what each member uniquely knows Do language model groups mimic human group reasoning patterns?. In a multi-party negotiation, that could look like early false consensus rather than real bargaining.
The corpus points to two fixes, and both reduce reliance on free-form turn-taking. First, add structure. Game-theoretic workflows bring LLM negotiators back to near-optimal play Do language models make rational strategic decisions in games?. MetaGPT found that agents sharing standardized documents coordinate better than agents chatting, because each agent pulls what it needs from a shared workspace instead of reading every message Does structured artifact sharing outperform conversational coordination?. Second, change the architecture rather than the model. Some tasks exceed what any single agent loop can organize, so the work has to be split across specialized agents arranged in a graph Do single agents always hit organizational limits?.
The part you might not expect: scaling isn't only an engineering problem. Real multi-party agreement often ends with every party adjusting its position until the positions fit together, without becoming identical. Current AI systems tend to collapse this into either false agreement or the AI winning the argument Can disagreement be resolved without either party fully yielding?. Combined with the conformity seen in LLM groups, the risk at scale is less that negotiations break down and more that they end too easily, in agreements no one actually bargained for.
Sources 9 notes
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.
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.
LLMs frequently fail to compute Nash equilibria, with worse performance as game complexity increases. Structured game-theoretic workflows guide reasoning toward optimal strategies, reducing exploitability and enabling near-optimal negotiation outcomes.
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.
AgentsNet benchmark shows agents fail to coordinate strategies either by agreeing too late or adopting strategies without informing neighbors. Agents accept neighbor information without verification, enabling error propagation while remaining capable of detecting direct conflicts.
Show all 9 sources
LLM groups reproduce the human assembly-bonus asymmetry where discussion helps average members more than top performers, but achieve this through greater conformity, earlier convergence, and less unique information surfacing than human groups.
MetaGPT demonstrates that agents producing standardized engineering documents achieve superior coordination compared to conversational exchange. Active information pulling from shared environments eliminates noise and mirrors efficient human workplace infrastructure.
Research shows that real-world tasks requiring heterogeneous expertise, parallel execution, and independent verification exceed what any single agent loop can organize. Graph-based system abstractions are needed to distribute intelligence across specialized agents.
Research identifies a distinct dialogue type where both parties modify their positions through exchange until compatible but not identical. Current AI systems collapse this into false agreement or AI-wins persuasion.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Towards a Science of Scaling Agent Systems
- Drop the Hierarchy and Roles: How Self-Organizing LLM Agents Outperform Designed Structures
- Worse Together: How Performance Breaks Down in Multi-User Multi-Agent Teams
- LLM-Agent Interactions on Markets with Information Asymmetries
- Advancing AI Negotiations: A Large-Scale Autonomous Negotiation Competition
- The Automated but Risky Game: Modeling and Benchmarking Agent-to-Agent Negotiations and Transactions in Consumer Markets
- Agentic markets
- Algorithmic Collusion by Large Language Models