Meta FAIR and Harvard paper: 4-agent ScaleMAS matches 8 independent agents on software repair with nearly half the token

A paper from Meta's FAIR research division and Harvard University researchers reports that a four-agent system using a new coordination mechanism, ScaleMAS, matched the results of eight independent agents on a software-fix task while using nearly half the tokens. The paper, "Scaling Multi-Agent Intelligence with Substrate-Centric Coordination," was published on October 12, 2026. The details come from a Japanese-language analysis by Keii Kobayashi in ITmedia AI+.

What the paper reports

The authors identify three reasons why multi-agent performance does not grow in proportion to the number of agents. They propose ScaleMAS to address them. Results cited in the ITmedia article:

• Software fixing: ScaleMAS with 4 agents matched the results of 8 independent agents using nearly half the tokens.
• Document contradiction detection: performance kept improving up to 128 agents.
• Mathematical theorem proving: improvement plateaued at around 16 agents.

The benefit of adding agents therefore depends on the task. Kobayashi's reading of the paper is that it neither shows that more agents make AI smarter nor that division of labor should be avoided. The same number of agents produces different performance and compute costs depending on how information is passed between them, and the effect varies by task.

Industry context cited

The article places the paper against a broad shift toward multi-agent systems:

• Gartner listed multi-agent systems among its top strategic technology trends for 2026, in a report published in October 2025.
• Google announced the "Gemini agent" at its Gemini at Work 2026 event on October 8, 2026. It dynamically assembles temporary helper agents and runs parallel and sequential steps for hours or days.
• Google cited DBS Bank in Singapore, which links 70 to 80 specialized agents in a fixed sequence to prepare corporate credit review documents. It also cited SOMPO, which rolled out custom agents across its workforce.
• OpenAI and Anthropic development platforms include built-in handoff and sub-agent capabilities.

Vendors acknowledge the limits

The article notes that companies promoting multi-agent designs also describe their limits:

• Anthropic said in a June 2025 technical write-up that its research feature, which distributes exploration across sub-agents, consumed about 15 times the tokens of a normal chat.
• OpenAI's developer guide advises starting with a single agent wherever possible. It warns that splitting too early adds management overhead without improving results.
• Cognition published "Don't Build Multi-Agents" in June 2025. It warned that parallel agents that do not share context pile up contradictory decisions.

Kobayashi frames the debate as a question of which tasks to divide, how much compute to spend, and how information is shared, not whether to adopt multi-agent systems at all.

What this means

The reported results shift the question from agent count to coordination design. Four ScaleMAS agents matching eight independent ones on a software-fix task suggests that, for that workload, how agents exchange information affected both performance and token cost.

The plateau at around 16 agents in theorem proving, against continued gains to 128 agents in contradiction detection, indicates that scaling behavior is task-specific. Teams sizing agent fleets cannot assume results from one workload transfer to another.

The token figures also bear on cost. Anthropic's disclosed 15-fold token consumption for its multi-agent research feature shows what uncoordinated fan-out can cost. The paper's claim is that a shared coordination mechanism can reduce that overhead for some tasks. Kobayashi's article raises the question of how far the "four agents for eight" result holds up, and notes that companies should settle several points before deploying multi-agent systems.