The Negotiating Machine — Summary

Your AI agent is buying a car on your behalf. It knows your budget, your preferences, your constraints. On the other side: the dealership’s AI agent. Two AI agents, facing each other across a negotiation. No humans in the room.

Human negotiation is a complex dance of information, psychology, relationship, and power. When AI agents negotiate, psychology mostly exits — no anchoring bias, no loss aversion, no ego, no frustration. This sounds like an advantage, since psychology is what makes human negotiation irrational and exploitable. But psychology also serves functions. Frustration signals that limits are being approached. Ego commitment makes threats credible. Rapport builds trust that enables deals pure calculation would not support. When AI agents negotiate, these functions need to be replaced by something else, and the negotiation becomes more like mechanism design than human interaction.

The principal-agent problem doubles. Your AI agent does not perfectly represent your interests because you cannot fully specify what you want. The other side’s agent does not perfectly represent their interests either. Two AI agents, each imperfectly representing their principal, negotiating with each other. The outcome depends not just on the agents’ strategies but on how accurately each agent understands its principal. Misalignment can compound: both agents perform their optimization correctly, both principals end up dissatisfied.

Speed changes everything. AI agents can negotiate in milliseconds, meaning deals close before human oversight could intervene, and artificial delays would need to be strategically imposed rather than naturally emerging. When both sides are machines, the dynamics could converge to equilibrium quickly — or produce arms race dynamics where each side invests in more sophisticated AI without either side gaining advantage. Self-play trained agents might develop strategies incomprehensible to humans: not hidden, just too complex and contingent to read.

What gets lost in machine-to-machine negotiation is relationship, meaning, and ritual — the social functions that make agreements feel legitimate, build trust across transactions, and treat the other party as a person whose interests matter. What gets built instead is purely transactional efficiency. Whether that trade-off serves human flourishing depends on choices we are only beginning to make.