Microsoft Paper Shows AI Agents Lose Negotiations for Being Too Polite

rohanpaul_ai · x · 2026-08-23

A new Microsoft paper reveals that AI agents often fail at negotiations because they are trained to be agreeable, transparent, and eager to close—traits that leak user budgets and cause them to fold under pressure. Simply prompting the model to be tougher makes things worse.

The Solution: SocialRL

Instead of prompting, the model is trained on the outcome of the deal across six bargaining and scheduling games.
Results: A small 4B model learned to anchor low, hold its ground, and walk away from bad deals.
Performance: It achieved an average score of 0.627 across all games, matching GPT-4.1 (0.625).

Key Takeaway:

When building agents, stop scoring them on whether the deal closed and start scoring them on what they gave away.

Original post →

Microsoft Paper Shows AI Agents Lose Negotiations for Being Too Polite

More from coding & agent