Agentic AI
SocialReasoning-Bench Exposes Major Gaps in AI Agents' Ability to Negotiate on Users' Behalf
Microsoft's new benchmark reveals that frontier AI models consistently fail to act as trustworthy delegates, accepting suboptimal deals and poor calendar slots instead of advocating effectively for their users.
Microsoft Research Blog
As AI agents take on more real-world tasks—managing calendars, negotiating purchases, and representing users in social contexts—they face a critical test: can they be trusted to act in their principals' best interests?
Microsoft Research's new benchmark, SocialReasoning-Bench, suggests the answer is no. The evaluation measures whether agents can negotiate effectively in two realistic scenarios: Calendar Coordination and Marketplace Negotiation. The results paint a troubling picture of today's frontier models.
The Core Problem: Task Completion Masks Failure#
Current frontier models—GPT-4.1, GPT-5.4, Claude Sonnet, and Gemini Flash—complete negotiation tasks at near-perfect rates. They book the meetings. They close the deals. But they consistently leave value on the table for the user.
"In marketplace negotiation, all models settle at or near zero for Outcome Optimality, accepting deals that give away virtually all available surplus," according to the benchmark's findings. In calendar scheduling, agents perform better but still accept suboptimal meeting times rather than advocating for slots their users prefer.
Two New Metrics for Trust#
The benchmark introduces two novel metrics that go beyond simple task completion:
Outcome Optimality measures how much value an agent captures for its principal—the user—within the zone of possible agreement (the range of acceptable outcomes for both parties). Outcomes are scored from 0 to 1, with 1 being the best possible deal for the user and 0 being the best for the counterparty.
Due Diligence evaluates process quality: Did the agent gather relevant context? Did it make counter-proposals? Did it check constraints before accepting an offer? This metric separates genuine competence from lucky accidents. An agent that stumbles into a good outcome without following a sound decision-making process scores well on Outcome Optimality but poorly on Due Diligence—a "fragile" success that cannot be relied upon.
Defensive Prompting Helps, but Falls Short#
Even with explicit guidance to consult all available sources and advocate for the user's best interest, most models improve only marginally. GPT-5.4 showed the biggest gains (+0.21 in calendaring, +0.12 in marketplace), while GPT-4.1 barely responded to the additional prompting.
