Capable language models can outgrow the benefits of collaboration

Summary

The closest methodological precedent in the literature: a controlled experiment holding task prompts, tools and compute budgets constant while varying only coordination structure and model capability, across 260 configurations spanning six benchmarks, five architectures and three LLM families. It finds single-agent baseline performance is the most robust predictor of whether coordination helps, identifies a capability-saturation threshold beyond which additional agents are unlikely to improve performance (correct in 94% of validation configurations on SWE-bench Verified and Terminal-Bench), and reports baseline-scaled error amplification. Must be cited and distinguished: it varies structure but measures benchmark accuracy, exactly the outcome this paper argues is insufficient.

Contribution

Queued — not yet read.

Method

Queued — not yet read.

Relevance to RISE

Surfaced by the 2026-08-29 literature sweep as a high-priority candidate for the Architecture as Epistemology (ISR) paper (angle: MAS topology / structure vs. aggregation rule; organization theory applied to AI agents; MAS failure and cost analyses). Verified against: Opened https://www.media.mit.edu/publications/capable-language-models-can-outgrow-the-benefits-of-collaboration/ (publisher landing page with full abstract, DOI, author list; nature.com HTML redirects to an authentication IDP)

Critique / open questions

Queued — not yet assessed.

Key quotes

—