AI Personal Assistants Struggle with Real-World Complexity in New Benchmark

The pitch for AI personal assistants has always been the same: Give the agent access to your digital life and it handles the rest. Your emails, your calendar, your notes, your devices—all of it. Your AI knows. Your AI acts. You sleep. Researchers from Huawei Technologies, Beijing Institute of Technology, Peking University, and the Chinese Academy of Sciences just built a benchmark to see if that's actually true. Spoiler: It's not. Claw-Anything evaluates AI agents across three dimensions at once: long-horizon event streams covering more than three months of simulated user activity, interdependent backend services averaging 10.1 per task, and multi-device interaction across both CLI Linux environments and GUI Android environments. The average context window per task is 191,700 words. Most existing benchmarks sit somewhere between 1,700 and 12,000. That's not a small gap but an entirely different problem. It’s also what real life feels like, as opposed to standardized ultra specific benchmarks. Your AI has no idea what's going on The benchmark is scored on pass@1—the probability the agent completes a task correctly on its first try, no do-overs. A task might ask the agent to cross-reference a price alert on a product it found weeks ago, check the user's calendar for a relevant appointment, and act on both from a phone. Another might ask it to pull recent work from notes, email threads, and Slack, then produce a presentation from scratch. These are things people actually ask assistants to do. Turns out AI isn't very good at them. GPT-5.5, per Decrypt's previous coverage, is OpenAI's best model, built with agentic, long-horizon tasks in mind. It scored 34.5%. "Current models remain unreliable even when given broader access to the user's digital world," the Claw-Anything paper reads. Several models that look impressive on other benchmarks dropped further. The benchmark also grades proactive assistance separately, meaning cases where the agent spots a need and acts without being asked. Most benchmarks don't test this. Claw-Anything does, and the gap is stark: Agents scored 25.9% on reactive tasks and just 6.7% on proactive ones. Why most benchmarks don't tell you this The researchers make a pointed argument: Existing benchmarks treat AI agents like task solvers given a clean desk. Claw-Anything treats them like personal assistants dropped into an actual messy life—irrelevant events, conflicting signals, months of accumulated noise. The agent has to figure out what's relevant before it can do anything useful. The ablation results make the multi-service dependency especially clear. When tools required for cross-service tasks were removed, success rates fell to nearly zero, because most tasks require agents to retrieve information and act across multiple backends rather than within a single one. This is not a new genre of problem in AI evaluation. OpenAI declared SWE-bench contaminated earlier this year after scores collapsed from roughly 70% to 23% on a less leakage-prone version. That was about data hygiene. This is about something more fundamental—whether the benchmarks are even asking the right question. On the constructive side, the team released the pipeline that generated the benchmark alongside 2,000 training environments. Fine-tuning Qwen3.5-27B on 1,500 successful agent trajectories improved pass@1 by 23.7%—enough to beat several closed-source models on the leaderboard, including Claude Sonnet.