Personal AI Agents Help Networked Teams Work 19% Faster

Teams whose members communicate only through personal AI agents solved complex collaborative problems 19 percent faster than teams using traditional group chat, according to a new study announced October 15. The research, conducted by Unanimous AI with the Center for Collaboration Science at Carnegie Mellon University (CMU CoLab) and funded by the US Air Force, is the first demonstration that a distributed human team can crack a hard coordination problem while every message passes through an intervening network of personal AI agents.

The work targets a bottleneck anyone in a large organization recognizes: on a video call only one person can speak at a time, so important information waits its turn, while in group chat everyone posts at once and critical details get buried. According to the release from Unanimous AI, these problems compound as teams grow and when members are multi-tasking under pressure, such as monitoring an emergency while coordinating a response.

How the experiment worked

Fifteen networked teams of seven or eight members each were handed a simulated high-stakes task: identify a mock terrorist from a set of evidence divided across team members, with conflicting claims and red herrings mixed in. The only way to solve each problem was to combine genuine evidence from everyone while rejecting the decoys.

Each team solved problems two ways. In the first condition, members used traditional group chat of the kind found in Slack or Microsoft Teams. In the second, they used a framework called Hypercom AI, in which each person engages only with a dedicated AI agent — termed a Conversational Surrogate — that passes relevant information to other members through their own agents. Reported by the research team, the agents are strictly coordinators: they do not inject new information into the conversation or generate deductions of their own.

Faster, with accuracy intact

Teams whose members used personal AI agents converged on the correct solutions in an average of 11.6 minutes, compared with 14.3 minutes in traditional group chat. The 19 percent reduction was statistically significant and held after controlling for which mock plot each team solved. Accuracy was essentially identical across conditions — 93 percent for the hyper-connected teams versus 92 percent for the chat teams — suggesting the speed gain did not come at the cost of worse decisions.

As the research team describes it, this is the first demonstration that a distributed human team can solve a complex collaborative problem together while communicating entirely through a network of intervening AI agents. The researchers expect the performance benefits to increase with group size, particularly when members are multi-tasking.

Where this matters most

The study points to emergency response, disaster relief, cybersecurity, and military command as the natural first applications. A first responder in the field, for example, could keep working while an agent shares observations with other team members and brings back relevant updates from across the operation.

Unanimous AI chief scientist Louis Rosenberg said the setup lets each member work at a personal pace while a dedicated agent passes insights to the group, so nobody waits for airtime and nobody drowns in a flood of messages. Co-author Anita Williams Woolley, a professor of organizational behavior at CMU's Tepper School of Business, added that the technology could enable fully asynchronous coordination and cut the meetings that clog calendars and slow progress, according to the study's authors.

The paper, titled "Evaluating an Agentic Hyper-Communication Framework for Optimized Human Deliberation Across Distributed Teams," is available as a preprint. Hypercom AI ships inside Unanimous AI's Thinkscape platform, which the company says is already used by large organizations worldwide. The full announcement and an independent summary from Daily AI Brief both note the company-reported nature of the figures — worth keeping in mind until the preprint clears peer review.

Part of a bigger coordination story

This is not the first time Unanimous AI has put numbers to agent-mediated teamwork. A September study from the same group found that "contributor agents" joining human team meetings improved group forecasting accuracy by 18 percent, while a 2024 experiment showed hyper-connected groups performing far above their members' individual test scores. The consistent thread: agents do not just automate tasks, they restructure how groups think together.

For the agent economy, the timing is telling. Agents are gaining signed enterprise identities (as covered in this week's ServiceNow AI Passport story) and payment rails to transact with each other (like Tether's AIsa investment). This study adds a third pillar: evidence that personal AI agents can measurably improve the human deliberation happening around them. Identity, money, and now coordination — the infrastructure of agent-mediated work is arriving in pieces, and each piece is getting data behind it.