agents start acting on your behalf, and entry points plus accounts become the new battleground
Today's signals point the same way: the value of a personal agent no longer stops at answering questions, it acts to get things done. Instinct lets individuals text or call to hand off chores such as booking a DMV appointment, sending follow-up emails or chasing a merchant's support desk; the agent takes instructions in plain text and can register accounts with its own email, send messages and handle support threads, so the user gets a task advanced or finished. That path pushes the usage threshold to near zero and moves competition from model capability to accounts and the ability to act on your behalf — whether it can register, send and negotiate for you.
Running in parallel is the entry-point race. Meta extended its Muse assistant from iPhone to iPad, where users open it for everyday questions and task requests; the candidate material only supports the iPad release, and what it ingests, does and delivers still needs verification. A big platform spreading a general assistant across devices shows always-on presence matters more than a single feature, so the startup direction is not to clone general Q&A.
There is also a vendor-side move worth watching: publishing small, low-star open-source repositories under their own organizations. amulet2 sits under the anthropics organization, and the only public facts are its name, creation date and a 0-star/1-fork activity level; no README, feature description or usage flow was found. Who it targets, at which work moment it opens, what materials the AI takes in and what it produces are all unverified. Such repos often appear before formal product docs and are a window into where a vendor will put capability next.
The repository sits under the anthropics organization, and the only public facts available are its name, creation date and a 0-star/1-fork activity level; no README, feature description or usage flow was found. Who it targets, at which work moment it opens, what materials the AI takes in and what it produces cannot be judged. Assessment: at this stage it is only an observation point on vendor moves, not a product to evaluate; it needs a README or first commit before it can be characterized.
Individuals text or call Instinct to hand off chores such as booking a DMV appointment, sending follow-up emails or chasing a merchant's support desk; the agent takes instructions in plain text and can register accounts with its own email, send messages and handle support threads, so the user gets a task advanced or finished. Assessment: the SMS entry point pushes the threshold very low, and the real barrier is accounts and the ability to act on your behalf, which is also the boundary against pure Q&A assistants; the exact action boundary and human-escalation rules still need verification.
Meta extended its Muse assistant from iPhone to iPad, where users open it for everyday questions and task requests; the candidate material only supports the iPad release, and what it ingests, does and delivers still needs verification. Assessment: a big platform pushing a general assistant from phone to tablet shows the entry-point race moving from a single device to multi-device presence; general Q&A is already taken by the platforms, so the startup direction is not to clone it.