On-Device AI Agent
An on-device AI agent runs locally on an end-user device, such as a smartphone, leveraging on-device inference capabilities to complete tasks without uploading data to external servers. This architecture provides privacy-protection advantages, but introduces distinct challenges around authentication, permission management, and security when the agent needs to call external APIs.
Timeline (last 90 days)
StareBrain splits the context needed by an on-device AI Agent into two categories: action context (describing the user's requested action, its parameters and the App that can complete it) and device context (information the phone already knows, such as upcoming calendar events, saved locations and recent search history).
On-device Agents are constrained by local compute and memory, with limits on model size, context length and inference speed, so they cannot simply copy the cloud paradigm of "a large model reads all the context and then plans," and instead must break tasks into smaller, more deterministic steps.
Most current mainstream AI agent products put reasoning and execution in the cloud, issuing instructions via phone or web page, with requests sent to remote servers to complete reasoning and then pushed back to the terminal.
All Facts (3)
StareBrain splits the context needed by an on-device AI Agent into two categories: action context (describing the user's requested action, its parameters and the App that can complete it) and device context (information the phone already knows, such as upcoming calendar events, saved locations and recent search history).
50%UnverifiedOn-device Agents are constrained by local compute and memory, with limits on model size, context length and inference speed, so they cannot simply copy the cloud paradigm of "a large model reads all the context and then plans," and instead must break tasks into smaller, more deterministic steps.
50%UnverifiedMost current mainstream AI agent products put reasoning and execution in the cloud, issuing instructions via phone or web page, with requests sent to remote servers to complete reasoning and then pushed back to the terminal.
50%