It runs up to 1.5x faster on long-horizon tasks.
Computer use agents are slow and brittle. The fix isn’t just stronger models, but also deploying them as multi-agent systems.
MACU is a general Multi-Agent Computer Use framework that consistently lifts success rates by 3.4-25.5% and is up to 1.5x faster on long-horizon tasks.🧵
This project actually started with JY creating a fake slack of our lab and the "JY" agent asking the "Lawrence" agent to send the "Russ" agent edited memes
I though that was honestly much better than running multi-agent computer use on academic benchmarks out of status quo, but I am pretty excited to work on the direction of multi-agent CUAs in the coming months - great work from Mr. Koh
Computer use agents are slow and brittle. The fix isn’t just stronger models, but also deploying them as multi-agent systems.
MACU is a general Multi-Agent Computer Use framework that consistently lifts success rates by 3.4-25.5% and is up to 1.5x faster on long-horizon tasks.🧵
MACU achieves better scaling behavior than single-agent CUAs, and improves success rates consistently across four CUA benchmarks (+4.7% on OSWorld, +3.4% on Online-M2W, +8.7% on WebTailBench, +25.5% on Odysseys).
MACU also reduces the wall-clock time on OSWorld and Odysseys.
MACU is simple and general: a manager decomposes tasks into a directed acyclic graph (DAG), dispatches parallel subagents, and revises the DAG with new findings.
A single slow CUA → a team of CUAs working in parallel!
Interactive visualizations + code: http://jykoh.com/macu
Read our paper to find out which features are important for building MACU systems. We find that the most important knobs are: (1) how often the manager is allowed to replan (2) the number of parallel subagents (3) the capability level of the CUA subagents and manager models.
We also find that MACU has better scaling trends than even pass@k (which cheats by using the groundtruth evaluator), making MACU a great replacement for tree search, best@k and pass@k frameworks!
MACU achieves better scaling behavior than single-agent CUAs, and improves success rates consistently across four CUA benchmarks (+4.7% on OSWorld, +3.4% on Online-M2W, +8.7% on WebTailBench, +25.5% on Odysseys).
MACU also reduces the wall-clock time on OSWorld and Odysseys.
It runs up to 1.5x faster on long-horizon tasks.
Computer use agents are slow and brittle. The fix isn’t just stronger models, but also deploying them as multi-agent systems.
MACU is a general Multi-Agent Computer Use framework that consistently lifts success rates by 3.4-25.5% and is up to 1.5x faster on long-horizon tasks.🧵
Users praise the MACU Multi-Agent Framework for showing practical real-world agent deployment instead of just theory.
