Choose Your Agent: Tradeoffs in Adopting AI Advisors, Coaches, and Delegates in Multi-Party Negotiation
Citation. Zhu, K., Thain, N., Tsai, V., Wexler, J., & Qian, C. (2026). Choose Your Agent: Tradeoffs in Adopting AI Advisors, Coaches, and Delegates in Multi-Party Negotiation. arXiv. https://arxiv.org/abs/2602.12089
Summary. Zhu and colleagues run an online behavioral experiment with 243 participants playing repeated three-person bargaining games with one of three LLM assistance modalities: proactive Advisor recommendations, reactive Coach feedback, or autonomous Delegate execution. Holding underlying model capability constant, they find that participants preferred the higher-control Advisor interface, while the Delegate condition produced the strongest group-welfare gains and positive spillover patterns. The paper is useful for the wiki because it separates technical agent capability from realized welfare under endogenous adoption, interface friction, and user preference for control.
Key claims
- In a within-subject experiment with 243 participants across 81 three-person bargaining groups, access to a Delegate agent increased group-level scaled surplus relative to a human baseline (β = 0.084, SE = 0.040, p = .033/.034; Holm-Bonferroni adjusted p = .10), while Advisor and Coach access did not show comparable improvements.
- Individual-level surplus showed the same positive direction for Delegate access (β = 0.028) but did not reach corrected statistical significance, which the authors attribute partly to heterogeneous adoption rates and variance in individual outcomes.
- Non-users in Delegate-available groups achieved a mean individual scaled surplus of 0.218, 21.6% higher than the human baseline mean of 0.179, though the regression result only approached raw significance (β = 0.039, SE = 0.020, p = .054; adjusted p = .16).
- Participants preferred the Advisor modality despite stronger welfare results for Delegate access, showing a preference-performance misalignment around control, cognitive effort, and willingness to let an agent execute decisions.
- Offer-generation assistance was used more than response assistance across all modalities, and response-side usage declined over negotiation rounds; requesting assistance did not guarantee adoption except in Delegate mode, where acceptance was structurally forced by autonomous execution.
- The authors interpret Delegate gains as a market-making mechanism: autonomous agents inject more rational, Pareto-improving proposals into the bargaining environment, improving the opportunity structure even for participants who do not directly use the agent.
Evidence & limitations
- The design cleanly varies Advisor, Coach, and Delegate interfaces while holding the underlying LLM scaffolding constant, making the paper stronger for modality comparisons than broad claims about LLM capability.
- The setting is a stylized three-player bargaining game with performance-based incentives, not an organizational field deployment; external validity to workplace delegation, team coordination, and accountable professional decisions remains open.
- Several key effects are directionally meaningful but weakened by conservative multiple-comparison corrections, so the source should be treated as early behavioral evidence rather than settled proof that autonomous delegation reliably improves group welfare.
- The placeholder ACM DOI in the PDF is not real provenance; the durable public record is the stable arXiv page and arXiv DOI.