In a design write-up, Alibaba's Qoder team argues that the 'more proactive' label attached to personal agent products hides three separate requirements. The first is an independent identity: its own email account, phone number, wallet and machine, plus the ability to run scheduled or event-triggered automations and retain memory across sessions. The second is cross-app execution through connectors and an isolated environment, checking back with the user when state changes or authorisation is needed. The third is structured long-term memory that turns scattered signals from calendar, mail, notes and files into next steps. The article lists the failure modes when one piece is missing: identity without data sources leaves the agent with nothing to do, data without memory turns every notification into a broadcast, and memory without identity means the agent waits until the user opens a chat. Its account of where initiative comes from is a gap calculation, goal minus current state, which becomes a problem and then a task. Task-level agents already run that loop inside a single task; long-horizon agents extend it across a project or a personal objective, with the user absent most of the time and authorisation granted per subtask rather than once. Detection is split between polling on a schedule and subscription to changes as events, such as a merged pull request, a mention in a group chat or a calendar edit. The piece closes on defining goals and acceptance criteria up front as the way to bound execution risk.
Our takeThe useful reframing is that a proactive agent needs a measurable gap; with no stated goal the gap is undefined, and initiative degrades into notification noise no matter how good the memory is. It also explains why per-subtask authorisation rather than one blanket approval is the natural control point: the agent decides what to do, the user decides how far. The open question the framework leaves is how acceptance criteria get written for goals a user cannot fully articulate at the outset.