MyPCBench: A Benchmark for Personally Intelligent Computer-Use Agents
2026-06-15 • Machine Learning
Machine LearningComputation and Language
AI summaryⓘ
The authors created MyPCBench, a testing environment where computer assistant programs act like personal helpers on a simulated Linux desktop with real web apps and user data. They designed 184 tasks based on real user requests to see how well these assistants handle complex, personalized jobs. Testing six different assistant models showed that the best one completed just over half the tasks, with struggles mostly in tasks involving many apps or long steps. This work helps measure how well AI assistants can manage real, personal computer tasks. The authors made the environment and tests publicly available for others to use.
personal assistantLinux desktopweb applicationsAI benchmarkingtask automationcomputer-use agentspersonalizationClaude Opus 4.6OpenClaw communityagent evaluation
Authors
Lawrence Keunho Jang, Andrew Keunwoo Jang, Jing Yu Koh, Ruslan Salakhutdinov
Abstract
Current benchmarks for computer-use agents evaluate models in impersonal environments. This leaves a gap between evaluation and deployment where personal assistants are expected to work across a user's whole digital life, including their context, historical data, and logged-in accounts. This gap is widest on web tasks, where live web evaluations cannot exercise sites that require logging in or personal information, the kind of site a real personal assistant has to drive. We introduce MyPCBench, which tests computer-use agents as personal assistants on a Linux desktop populated with 17 simulated real-world web applications and a full desktop stack, all seeded for one canonical persona, Michael Scott from The Office. We define 184 tasks in this environment, each inspired by a real request drawn from the OpenClaw community, and benchmark six closed and open-weight models with a uniform computer+bash tool surface. We find that the best model, Claude Opus 4.6, fully solves 55.4\% of the tasks, the only model above 50\%. Model failures cluster on tasks that span many applications and on long trajectories, where personalization stresses an assistant the most. We release the environment, task set, and agent harness at https://mypcbench.com.