Do Multi-Agents Dream of Electric Screens? Achieving Perfect Accuracy on AndroidWorld Through Task Decomposition

Pierre-Louis Favreau, Jean-Pierre Lo, Clement Guiguet, Charles Simon-Meunier, Nicolas Dehandschoewercker, Allen G. Roush, Judah Goldfeder, Ravid Shwartz-Ziv · 2026

Read Do Multi-Agents Dream of Electric Screens? Achieving Perfect Accuracy on AndroidWorld Through Task Decomposition by Pierre-Louis Favreau, Jean-Pierre Lo, Clement Guiguet, Charles Simon-Meunier, Nicolas Dehandschoewercker, Allen G. Roush, Judah Goldfeder, Ravid Shwartz-Ziv (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-07787.txt.

Description

Phone-use benchmarks test whether AI agents can operate Android apps, and until now no system had solved all of them. Pierre-Louis Favreau and seven co-authors present Minitap, a multi-agent system that scores 100 percent on AndroidWorld's 116 tasks, above the 80 percent human level.

Single agents fail, the authors show, because mixed reasoning traces pollute context, silent text-entry failures go unnoticed, and action loops repeat without escape. Minitap answers with six specialized agents, deterministic checks that typed text actually landed, and meta-cognitive cycle detection. Ablations credit decomposition with 21 points, verified execution with 7, and meta-cognition with 9. The open-source system sets a new bar for mobile agents.

Catalog record

CREATORPierre-Louis Favreau, Jean-Pierre Lo, Clement Guiguet, Charles Simon-Meunier, Nicolas Dehandschoewercker, Allen G. Roush, Judah Goldfeder, Ravid Shwartz-Ziv
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE39 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-07787

Browse the full library · About the Tower of Babel