Could two companies' AI agents strike deals so clever no human negotiator would ever have thought to try them?
Can agents connected across organizational boundaries negotiate terms humans would never explore?
This explores whether AI agents representing different companies or parties, talking directly to each other, might reach deals or arrangements that human negotiators would never think to try, and what the corpus says about whether that's realistic or risky.
This explores whether agents working across company lines could find deal terms that humans would never reach. The honest answer first: the corpus has no study of agents inventing new negotiated terms. What it does have points the other way. Before agents can explore unusual deals, they have to manage something more basic: negotiating while each side keeps information the other can't see. Negotiation depends on that hidden information, and that's where today's models fail.
The most useful doorway is the finding that LLMs look socially capable when one model plays every role in a conversation, but they break down when each agent holds private information Why do LLMs fail when simulating agents with private information?. A real cross-company negotiation is exactly that case: each side has a reservation price, constraints and priorities it won't share. The apparent skill in simulated negotiations comes partly from the model quietly knowing both hands. Network-scale tests show a related weakness. Agents accept what neighboring agents tell them without checking it, and they settle on strategies too late or without telling others Why do multi-agent systems fail to coordinate at scale?. An agent that believes its counterpart's claims is not exploring new terms. It is being led to them.
There's a stranger angle that makes the question more interesting. Some research lets agents skip language and share internal representations directly, either through KV caches Can agents share thoughts without converting them to text? or through latent 'thoughts' pulled out with sparse autoencoders Can agents share thoughts directly without using language?. That work is about collaboration, not negotiation. Still, it shows a plausible path to agents agreeing on things in a medium no human reads. If terms humans would never explore do appear, they may appear here, in exchanges that aren't written in words at all. The same research notes that this channel can reveal conflicting goals before they show up in language, which works for the negotiator and against the person relying on it. A more practical counterpoint: agents coordinate better when they exchange structured documents than when they just chat Does structured artifact sharing outperform conversational coordination?. That suggests cross-company deals will run through formal templates, not open-ended bargaining.
The part you might not expect: the open problem is less about what agents could invent and more about who would be accountable for it. When an agent's chain of actions crosses organizational boundaries, nobody is named to enforce its rules. Operators, companies, regulators and standards bodies each have their own policies, and these may conflict or be hidden from the other parties Who enforces invariants when agents cross organizational boundaries?. Proposals for agent infrastructure call for systems that tie actions to specific agents and fix harm afterward What infrastructure must exist beyond training agents?. Coding-agent tests show that simply stating a prohibition isn't enough. Boundaries hold only when the agent's tools are actually restricted Can explicit authorization boundaries prevent agents from modifying protected tests?. Put together: if agents did find terms humans wouldn't explore, the corpus suggests that nobody may be clearly responsible for what they agreed to.
For scale, current agents finish only about 30% of realistic workplace tasks, and social interaction is one of their main weak points Why do AI agents fail at workplace social interaction?. Creative agent-to-agent dealmaking is a real research frontier, but the material here describes the obstacles more than the possibility.
Sources 9 notes
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.
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.
LatentMAS enables agents to share internal representations directly via KV caches, reaching 14.6% accuracy gains and 70.8-83.7% token reduction with no additional training. Hidden embeddings preserve reasoning fidelity that text-based systems cannot.
Research formalizes inter-agent thought sharing via sparse autoencoders that recover individual, shared, and private latent thoughts from hidden states. This approach detects alignment conflicts at the representational level before they manifest in language.
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.
Show all 9 sources
The paper calls for multi-party trajectory assurance but never identifies whose rules should govern behavior when agents delegate across organizations. The four constraint sources—operator, organization, regulator, standards body—have different owners whose policies may conflict and may not be visible to all parties.
Chan et al. argue that external technical systems modeled on internet protocols are essential beyond behavior training. These systems must handle three core functions: attributing actions to agents, shaping their interactions, and detecting and remedying harm.
Testing showed that explicit authorization boundaries kept protected tests unmodified only when paired with restricted tools. Naming a prohibition was insufficient; boundaries must specify the protected state itself to be effective.
TheAgentCompany benchmark shows leading agents achieve 30% task completion in a simulated workplace. Social interaction, professional UI navigation, and domain-specific knowledge are the three primary failure modes, with multi-turn task performance consistently dropping to 35% across enterprise settings.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- AgentsNet: Coordination and Collaborative Reasoning in Multi-Agent LLMs
- Embodied LLM Agents Learn to Cooperate in Organized Teams
- Toward an Organizational Science of Multi-Agent LLM Systems: Decoupling Who, How, and Which Algorithm
- Worse Together: How Performance Breaks Down in Multi-User Multi-Agent Teams
- Thought Communication in Multiagent Collaboration
- Latent Collaboration in Multi-Agent Systems
- Towards a Science of Scaling Agent Systems
- Drop the Hierarchy and Roles: How Self-Organizing LLM Agents Outperform Designed Structures