What happened
Decrypt reports that Huawei Technologies, Beijing Institute of Technology, Peking University, and the Chinese Academy of Sciences released Claw-Anything, a benchmark for personal-assistant agents operating inside a simulated digital life: email, calendar, notes, devices, and the messy context around them.
The reported result is the useful part. GPT-5.5 scored 34.5% on pass@1, even though current agent marketing often implies this category is nearly ready for broad delegation. The team also released an automated data pipeline that produced 2,000 training environments, and fine-tuning an open-weight model on that data improved task success by 23.7%.
This is the signal builders should care about: agents will not be judged by how well they answer tidy prompts. They will be judged by whether they can maintain context, recover from ambiguity, and avoid expensive mistakes in real workflows.
The boardroom question is shifting from "which model is smartest?" to "which operating environment makes delegation reliable enough to trust?"
Source
Reported by Huawei's New Benchmark Gives AI Agents Months of Your Life-Then Watches Them Fail via decrypt.co, published May 27, 2026.