Oxford study shows that AI agents do not work with new partners
Written rules in English helped the agents to select safe steps, but the tests were small.
Researchers at Oxford University put a paper on arXiv on 4 November 2026. They say that top language model agents do not work together with partners that they do not know. The team made small tests for this task. The report gives no test numbers. The team says that written rules help the agents to select safe steps.
How it works
The team works on zero shot coordination. Agents must work with a new partner and have no practice. The team gives each agent a short set of written rules. The rules push the agents to safe steps and make them do the same.
The proof
The agents did not do correctly in the small tests. The team says that the agents do not know how teamwork works. The team says that the rules gave a large increase, but the report gives no number. A computer method found rules that gave the same result as rules that persons wrote, or more.
What is not known
A warehouse robot or a large software project is much more than a small test. It is not known if the rules work in large tasks that change.
Insight: Agents from different builders have no shared rules
In October 2026, a test found that agents frequently do not release locks. A test of many agents found that they moved into opposite groups. The Oxford result adds that agents do not read new partners correctly. It seems that the first written rules in use will set the standard for all agents. Look for who writes them.
Sources
Posted
