π-Bench asks whether your agent can read the room

Most agent benchmarks test one thing: can you do the task I just told you to do. π-Bench, which picked up 75 upvotes on Hugging Face papers, tests the opposite and much harder thing: do you notice the task I did not tell you about. It is a benchmark for proactive personal assistants, a hundred multi-turn tasks across five different user personas, and the whole point is to measure whether an agent acts on an unstated need before being asked.

The findings are a useful reality check. Proactivity is genuinely hard for current agents. There is a wide, measurable gap between an agent that can complete a task and one that can anticipate it. And, most interestingly, prior interactions help a lot, an agent that remembers earlier turns and earlier sessions is much better at resolving what you meant but never said. In other words, memory and continuity are not nice-to-haves for proactivity, they are the mechanism.

Why this is worth your attention: the difference between a tool and an assistant is exactly this axis. A tool waits to be told. An assistant notices. Every product chasing the personal-agent dream, the OpenClaw-style always-on helper, is implicitly betting it can cross this gap, and until now there was no clean way to score who is closer. Paper at arxiv.org/abs/2605.14678.