It is a coordination system proposed in the Meta FAIR and Harvard paper "Scaling Multi-Agent Intelligence with Substrate-Centric Coordination" that addresses why multi-agent performance does not scale linearly with agent count, focusing on how information is shared between agents.

Did the study find that more AI agents always perform better?

No. Gains kept rising up to 128 agents for document contradiction detection but plateaued around 16 agents for math theorem proving, and on software repair 4 ScaleMAS agents matched 8 independent agents with nearly half the tokens.

What do OpenAI and Anthropic say about multi-agent setups?

OpenAI's developer guide advises starting with a single agent where possible, warning that splitting early adds management overhead without improvement. Anthropic said its research feature's token use was about 15 times that of normal chat.

Large Language ModelsAI Safety & AlignmentITmedia AI＋12h agoOct 29, 2026, 10:01 JST

ScaleMAS: 4 AI agents match 8 on half the tokens, Meta-Harvard study finds

Send on LINE

3 Key Points

What happened
A paper titled "Scaling Multi-Agent Intelligence with Substrate-Centric Coordination," published Oct 12 by Meta's FAIR lab and Harvard researchers, proposed ScaleMAS and reported that 4 ScaleMAS agents matched 8 independent agents on software-repair tasks using nearly half the tokens.
Why it matters
The findings suggest adding more AI agents does not automatically improve results — the coordination mechanism and the task type appear to matter more, which challenges the assumption that more agents is better.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Get these free every morning

FAQ

What is ScaleMAS?: It is a coordination system proposed in the Meta FAIR and Harvard paper "Scaling Multi-Agent Intelligence with Substrate-Centric Coordination" that addresses why multi-agent performance does not scale linearly with agent count, focusing on how information is shared between agents.
Did the study find that more AI agents always perform better?: No. Gains kept rising up to 128 agents for document contradiction detection but plateaued around 16 agents for math theorem proving, and on software repair 4 ScaleMAS agents matched 8 independent agents with nearly half the tokens.
What do OpenAI and Anthropic say about multi-agent setups?: OpenAI's developer guide advises starting with a single agent where possible, warning that splitting early adds management overhead without improvement. Anthropic said its research feature's token use was about 15 times that of normal chat.

ITmedia AI＋Read Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.