A team of AI agents had a lower score than one agent in 12 of 13 tests

The researchers did a test of one AI agent and a team of AI agents. The agents work with a user and follow written rules. The test used about 7,950 test conversations and 6 models. The team had a lower score in 12 of 13 test groups and was equal in the other. The average difference was 28 points.

How it works

The tests included airline, retail and telecom tasks. The team had a planner and used the same model as the one agent. A test user was the customer. The planner did not see how the tools work.

The numbers

In 2 groups, a rule from other tests says the team gets more points. There the team had 13 and 22 points less. With no budget limit, the team was only equal and used 5.5 times the tokens. It did not decrease incorrect writes to the database. It changed other errors into budget limits and no work.

What is not known

One increase was found, in the group near 45% precision for one agent, with a different test user. The researchers do not know if the increase gets less when models get stronger. The users in the tests were not persons.

Insight: Teams of agents win on tasks with no user

A report of 12 October 2026 showed that 2 agents got a medal on 78.2% of MLE-bench tasks. The agents work one after the other. The test has no user. This study finds the opposite with a user and written rules. This reads as two cases: agent teams do well on tasks with no user and do badly in customer work. Leaders must get test results from their tasks.