Meta and Harvard Researchers Study AI Teamwork

Researchers from Meta’s FAIR lab, Harvard University and other institutions published a paper proposing ScaleMAS on October 12, 2026. The method aims to improve coordination among multiple AI agents. The researchers identified three reasons performance may fail to rise in proportion to agent count and proposed a coordination approach centered on information sharing.

On software repair tasks, four ScaleMAS agents matched the performance of eight independent agents while using nearly half as many tokens. Performance on document contradiction detection continued to improve through 128 agents, but gains in mathematical theorem proving plateaued at about 16. The benefit of adding agents varied by task.

Dividing work among multiple agents is spreading among companies and AI developers. In October 2025, Gartner named multi-agent systems—where agents coordinate on complex goals—as one of its strategic technology trends for 2026. OpenAI and Anthropic developer platforms include features for handing work between agents and calling helper agents. At “Gemini at Work 2026” on October 8, 2026, Google announced Gemini agent, which can assemble helper agents as needed and run parallel and sequential workflows for hours or days.

In an example presented by Google, Singapore’s DBS Bank links 70–80 specialized agents in a prescribed workflow to prepare corporate credit review materials. SOMPO deployed more than 10,000 custom agents for its 34,000 employees. Anthropic said in June 2025 that a research feature distributing searches among multiple helper agents was effective, but used about 15 times as many tokens as a standard chat.

OpenAI’s developer guide recommends starting with one agent, saying that dividing work too early may add management overhead without improving performance. AI coding company Cognition also warned in a June 2025 article that parallel agents that do not share context can make conflicting decisions. The study shows that the performance and computing costs of dividing work depend on the task and how information is shared. The factors for companies considering adoption include which tasks to divide, how much computing capacity to use, and how agents will share information.

Researchers from Meta’s FAIR lab, Harvard University and other institutions published a paper proposing ScaleMAS on October 12, 2026. The method aims to improve coordination among multiple AI agents. The researchers identified three reasons performance may fail to rise in proportion to agent count and proposed a coordination approach centered on information sharing.

On software repair tasks, four ScaleMAS agents matched the performance of eight independent agents while using nearly half as many tokens. Performance on document contradiction detection continued to improve through 128 agents, but gains in mathematical theorem proving plateaued at about 16. The benefit of adding agents varied by task.

Dividing work among multiple agents is spreading among companies and AI developers. In October 2025, Gartner named multi-agent systems—where agents coordinate on complex goals—as one of its strategic technology trends for 2026. OpenAI and Anthropic developer platforms include features for handing work between agents and calling helper agents. At “Gemini at Work 2026” on October 8, 2026, Google announced Gemini agent, which can assemble helper agents as needed and run parallel and sequential workflows for hours or days.

In an example presented by Google, Singapore’s DBS Bank links 70–80 specialized agents in a prescribed workflow to prepare corporate credit review materials. SOMPO deployed more than 10,000 custom agents for its 34,000 employees. Anthropic said in June 2025 that a research feature distributing searches among multiple helper agents was effective, but used about 15 times as many tokens as a standard chat.

OpenAI’s developer guide recommends starting with one agent, saying that dividing work too early may add management overhead without improving performance. AI coding company Cognition also warned in a June 2025 article that parallel agents that do not share context can make conflicting decisions. The study shows that the performance and computing costs of dividing work depend on the task and how information is shared. The factors for companies considering adoption include which tasks to divide, how much computing capacity to use, and how agents will share information.