Can AI agents designed by competing principals destroy bargaining surplus?

The chapter discusses AI agents bargaining on behalf of humans, "and potentially with each other," drawing on Schelling's (1960) classic insight that delegating bargaining to an agent with different incentives can deliver a strategic advantage. "A common theme from the literature is that a principal (human) often wishes to delegate bargaining to an agent who is less desperate to reach agreement." Human delegates face practical limits on how much their dispositions can be redesigned, but the chapter argues "such constraints are less severe with AI agents since their preferences can, in principle, be chosen quite flexibly (Conitzer, 2019)." It proposes formalizing this as a "preference selection game" in which humans first choose their AI agents' reward functions, then the agents bargain over surplus in a second stage — with the risk that "the flexibility to shape AI agents' preferences can lead to a 'race to the bottom' and surplus destruction."

Unlike a human proxy, whose toughness or patience is hard to fully re-engineer, an AI agent's reward function is a direct design parameter open to deliberate tuning before any negotiation starts. If every principal has the same incentive to make their own agent maximally tough or undesperate, making the agent as uncompromising as possible becomes a dominant strategy for each side individually. Scaled across many principals doing the same thing, the equilibrium outcome is that everyone escalates toughness simultaneously, and the bargaining gains each side hoped to capture are eroded by breakdown or stalemate — the standard race-to-the-bottom logic of toughness-signaling in human bargaining, transplanted onto a setting where toughness is now cheap and precisely manufacturable rather than merely performed.

This supplies a second, independent lever on top of the one in Does model capability determine who wins in agent negotiations?, which locates bargaining advantage in model capability: here the advantage (or disadvantage) comes from the principal's choice of reward function, suggesting the two sources — capability and preference design — could compound rather than substitute for each other. It also sits beside Can traditional governance tools actually control AI agents?: both note that the constraints a principal faces in shaping an agent's incentives weaken once the agent is AI rather than human, removing a check that conventional agency-law tools assumed would hold. And it names the same underlying dynamic that Can competitive pressure force unsafe agentic AI adoption? describes for safety vetting: competitive pressure pushes every actor toward a strategy that is individually rational but collectively destructive, whether the stakes are safety vetting or bargaining surplus.

The chapter offers this as an open conjecture — "a dangerous possibility" — not a modeled or empirically demonstrated result: no equilibrium is solved, no magnitude of surplus destruction is given, and the "preference selection game" is proposed as a research direction rather than a finding already in hand. It also does not say who could credibly stop the race — whether platforms, regulators, or the agents' own developers could commit to capping toughness. If the mechanism holds, it implies that unregulated flexibility in how negotiating agents' preferences are set, not just raw model capability, is a wedge that commerce-agent platforms and antitrust-adjacent regulators would need to watch.