PalmClaw runs mobile agent loop on-device, cuts time 94.9%

The interesting bit in this month's mobile agent work is not a bigger model or a smarter planner, it is that the whole loop ran on the phone. In a paper posted to arXiv, Hongru Cai, Yongqi Li, Ran Wei and Wenjie Li describe PalmClaw, an open-source framework that keeps sessions, memory, skills, tools, and the agent loop resident on the device instead of shipping every step out to a server.

The reason that matters is the way most mobile agents work today. They drive the phone the way a human does, tapping, swiping, and typing through the GUI. The authors' framing is that this produces long, interface-dependent sequences that cannot reach device capabilities directly and make it hard to say where one action stops and the next begins. PalmClaw's approach is to expose those capabilities as tools with explicit arguments and structured results, so the agent calls the phone the way a program would rather than pantomiming a user.

The reported numbers are the part that will get attention: an 11.5% relative improvement in task success and a 94.9% reduction in completion time versus what the authors call the strongest baseline. If those hold up, the practical story is that a mobile agent does not need to round-trip to a hosted service for every step, which changes the cost, latency, and privacy picture for anyone building consumer agents on phones.

The honest caveat is that this is a single paper with the authors' own comparison. The abstract does not tell you which baseline system, which benchmark, what model is running on-device, or what the battery and memory cost looks like in a real session, and none of those are small details for a phone. Take the specifics as reported, not settled.

The direction is what is worth watching. If the on-device agent loop is competitive with a server-side one, the strategic question for handset makers, OS vendors, and consumer app teams stops being which cloud agent to integrate and becomes which device tools to expose, and to whom. That is a very different roadmap.