An agent phone is not useful until you can control every step it takes
STEPX Neo, Step AOS, and Amoo bring the agent-phone idea into focus. The test is not a smooth demo; it is whether cross-app permissions, approval for consequential actions, and recovery from mistakes are real.
Ask a phone to “order lunch for me” and it may simply open a delivery app and fill in a search box. That is a better shortcut. It starts to resemble an agent only when it can use your address, budget, dietary needs, and delivery deadline — then ask before placing the order. That is the line worth using to judge the new wave of agent phones.
On July 13, StepFun announced its STEPX device brand, the Step AOS operating system, the Amoo personal agent, and the STEPX Neo phone. Public reporting confirms the launch and how the products fit together. Labels such as “the world’s first,” however, should be read as company positioning, not as an independently established comparison.
Check whether the task closes, not whether the chat sounds natural
An on-phone agent needs to do four things: understand the request, obtain the needed data and permissions, act, and leave a result that can be reviewed. Miss any one of them and “cross-app work” can collapse into a demo with a few app hops.
Try a low-risk task first: ask it to compare two restaurants and prepare an order for approval, rather than placing a payment. Watch whether it identifies the apps it will access, the personal data it uses, the action it is about to take, and the way to cancel. Those four answers matter more than a claim that it can “get things done.”
An ecosystem is not an app list; it is a revocable permission boundary
StepFun says it is working with services including Alipay, Meituan, Amap, Didi, JD.com, and Baidu. A wider partnership list still does not mean an agent can perform every action in every service. Reading information, creating a draft, submitting an order, and making a payment carry very different risks.
Evaluate permissions separately. Can it search in read-only mode? Which step requires a fresh confirmation? Can access be withdrawn by app, data type, and duration? Is there a log that lets you reconstruct a bad action? An agent becomes a daily tool only when these controls are normal product surfaces, not footnotes.
Hybrid on-device and cloud processing can improve speed, not trust by itself
The phone can handle immediate and lightweight actions, while longer planning and complex reasoning may involve the cloud. That division can help responsiveness and capability. It does not answer when a request leaves the device, what is retained, or how a user erases memory.
Before buying or trying one, use a short checklist: Can sensitive data be excluded? Are payments, messages, permission changes, and deletion always confirmed? Can a failed task be returned to its prior state? Is the third-party permission scope explained? If any answer is vague, treat the feature as an experiment, not a dependable delegate.
The phone is still the entry point; service cooperation is now the exam
Detailed STEPX Neo hardware specifications and the real task catalog still need to be made public. The useful question today is not whether phones disappear. It is whether an agent can form a task contract between existing services that the user can understand and revoke.
Early users should begin with reviewable work: search, comparison, organization, and drafting. Keep payment, sending, permission changes, and deletion in human hands. The agent-phone contest will not be won only by who has the most fluent model; it will be won by who makes permission, confirmation, and accountability clearest.
Sources
Related
ChatGPT Work can connect apps and files, break down large tasks, and keep working for hours. Its real significance is that AI assistants now need the same permissions, boundaries, and acceptance criteria as coworkers.
PixVerse has shown a game-engine architecture that combines a real-time world model, game rules, and agent orchestration. It may reduce the cost of expressing a prototype, but it has not yet proved itself on latency, rule stability, or cost.