Why might AI agents bargain more patiently than humans do?

Shahidi, Rusak, Manning, Fradkin and Horton argue that AI agents "will also affect the prevalence and dynamics of bargaining as their opportunity cost is not human time." They contrast this with standard theory: "classic bargaining models assume impatience and a preference to conclude negotiations quickly because time and attention are scarce resources for humans." For an agent, they write, "the binding constraint is compute rather than time," so "even if the principal prefers timely resolution, agents can initiate negotiations earlier, hold many concurrent negotiations, and continue engaging in these for a longer period of time" — their example is an agent that could "begin negotiating summer-2027 rentals in January 2026."

The authors tie this to the same chapter's broader argument about agent pricing and supply: human agents are paid a share of transaction value partly because the best human negotiators are scarce and high-powered incentives buy extraordinary effort, whereas AI agents have "virtually unconstrained" supply and "do not derive utility from pecuniary compensation," which "point to diminished pricing power for AI relative to human agents." The counter-pressure they note is that if agent performance scales with compute allocated, prices "could scale with the amount of compute required," potentially rising with the stakes of the transaction. Patience in negotiation and compute-driven pricing are, in this account, two expressions of the same substitution: agents trade compute for the two things that disciplined human bargaining — time and money.

This complicates Does model capability determine who wins in agent negotiations?, which measures negotiation outcomes as a function of model capability inside a fixed-duration benchmark. If Shahidi and colleagues are right that deployed agents can hold many concurrent negotiations over extended, principal-chosen timeframes, compute-funded persistence becomes a second axis of advantage layered on top of the capability gap that benchmark finds — meaning unequal outcomes in agent-mediated bargaining could compound from both which model a party can access and how much sustained compute they can afford to spend holding out.

The claim is illustrated with a hypothetical (rental negotiations beginning over a year in advance) rather than measured negotiation behavior, and the chapter is a non-peer-reviewed NBER working paper chapter offering a conceptual argument, not an empirical test of how agents actually behave in live negotiations or how principals actually weigh speed against better terms. The implication the argument supports, at that strength, is that wherever a principal does want a fast resolution, there is a latent gap between what an agent is capable of (patient, parallel negotiation) and what it is instructed or incentivized to do — a gap the chapter identifies but does not resolve.