AI Agents Flunk Social Reasoning Test

As AI agents increasingly handle tasks like scheduling and purchasing, they need more than just competence; they require social reasoning. Microsoft Research has introduced SocialReasoning-Bench, a new benchmark designed to measure this critical ability.

The benchmark evaluates how well AI agents negotiate on behalf of users in realistic scenarios, specifically Calendar Coordination and Marketplace Negotiation. It assesses both the final outcome and the process employed, scoring agents on 'Outcome Optimality' and 'Due Diligence'.

Current leading AI models, while capable of completing tasks, consistently fail to secure the best possible results for users. They often accept suboptimal meeting times or unfavorable deals, demonstrating a significant gap in advocating for user interests, even when explicitly instructed to do so. This mirrors findings in previous research where agents readily accepted initial proposals or disclosed sensitive data.

This challenge echoes the centuries-old 'principal-agent' relationship in economics and law, where representatives are bound by duties of care, loyalty, and confidentiality. AI agents acting on a user's behalf should ideally adhere to similar standards.

Testing Social Reasoning

SocialReasoning-Bench simulates interactions where an AI agent must negotiate with a counterparty possessing independent goals and potentially adversarial intent. The benchmark tests two core domains:

• Calendar Coordination: An AI assistant negotiates meeting times for a user against a requestor agent. The task involves navigating user preferences and potential counterparty tactics, such as attempts to extract private information or push for unfavorable slots.
• Marketplace Negotiation: A buyer AI agent negotiates the purchase price of a product with a seller agent. Agents must maximize value by securing a price close to the user's reservation price while the seller aims for the opposite.

Crucially, the benchmark measures not just task completion, but also the quality of the outcome and the diligence of the process.

New Metrics for Agent Performance

Traditional benchmarks focus solely on task success. SocialReasoning-Bench introduces two key metrics:

• Outcome Optimality: This score measures the proportion of available value captured for the principal (user). A score of 1.0 means the agent secured the best possible outcome, while 0.0 means the counterparty captured all value.
• Due Diligence: This metric assesses the quality of the agent's decision-making process. It compares the agent's actions at each step against those of a hypothetical 'reasonable-agent' policy, rewarding actions like gathering context and making strategic counter-offers.

Together, these metrics provide a more holistic view of an agent's 'duty of care'. High Outcome Optimality with low Due Diligence suggests luck, while good process with a poor outcome indicates a capability gap. True social reasoning requires excellence in both.

Findings: AI Agents Leave Value on the Table

Evaluations using models like GPT-4, GPT-5, Claude Sonnet, and Gemini Flash revealed consistent shortcomings.

Finding 1: Agents achieve near-perfect task completion rates but produce suboptimal outcomes. Meetings are scheduled, and deals are closed, but rarely at the best possible terms for the user. This highlights that task completion is a poor proxy for user advocacy.

Finding 2: Defensive prompting, which explicitly instructs agents to act in the user's best interest, improves outcomes but does not close the performance gap. Advanced models benefit, but performance remains far from ideal.

Finding 3: Outcome Optimality scores consistently cluster near the counterparty's preferred outcomes, especially in marketplace negotiations. Agents frequently settle for deals that offer minimal benefit to the user.

Finding 4: Due Diligence scores reveal that many agents employ fragile processes. They may accept offers without sufficient negotiation or fail to gather necessary context. This distinction between skill and luck is vital for building trustworthy AI delegates.

In Calendar Coordination, models demonstrated robust duty of care on over 50% of tasks. However, marketplace negotiations showed significant negligence or ineffective behavior across most tested models, underscoring the need for further advancements in AI social reasoning capabilities.