← News

Agents & MCPSources: arXiv · 18 Nov 2026

Study: web research agents interpret the same priority with different criteria

A new arXiv paper reports that two leading web research agents followed a stated priority most of the time, yet judged companies by different standards, which shaped their conclusions.

In plain words

Web research agents are AI tools that search the internet and write reports with citations. Researchers asked two such agents to compare Microsoft and Alphabet while stressing either growth or resilience. The agents usually honored the stated priority, but each measured it in its own way, which changed the verdicts.

What was announced

The study, posted as arXiv:2611.22989, examined two agents built on OpenAI and Anthropic models. Across 36 runs, prompts asked the agents to compare Microsoft and Alphabet as businesses, with special attention to growth or resilience, or with no priority at all. The authors analyzed the reports, searches, returned sources and cited sources. Of the 24 reports written under a priority, 21 kept the focus on it, but the criteria behind that focus varied.

The differences mattered for outcomes. Under the growth priority, OpenAI-based reports that ranked the companies weighed risk or predictability and favored Microsoft, while Anthropic-based ones measured growth by its pace and favored Alphabet. Under resilience, the reports applied partly different criteria, and most named Microsoft as the sturdier business. Among the search and source measures tested, only the Anthropic agent's search wording changed detectably between conditions. The authors suggest that evaluations should look at the criteria behind a report's judgements, not just its focus or accuracy, and that interfaces could let users view and adjust those criteria.

What it does not show

  • The evidence is an abstract, so methods and results are available only in summary and have not been independently confirmed.
  • The study covers two agents, two companies and 36 runs, so the findings may not extend to other agents, prompts or tasks.
  • The paper reports which agent families behaved differently but does not explain why the criteria diverged.

For security and IT teams

Scope
The findings concern two research agents comparing Microsoft and Alphabet in a controlled study, not a general audit of any commercial product.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.