Org-Agent: Beyond Personal Assistants Towards Organizational Agents

Org-Agent. Beyond Personal Assistants Towards Organizational Agents. By Luyao Zhuang and four others, from the Hong Kong Polytechnic University. Published on the twenty-eighth of September. No code repository is linked to it.

An agent serving an organisation, rather than one person, must coordinate requests from multiple users while using knowledge spread across all their interactions. The authors identify two complementary capabilities this calls for. One is interaction and decision-making across users. The other is memory and knowledge use across users. The first is about acting on behalf of several people at once. The second is about remembering what each of them said, and using it correctly later.

Both are governed by organisational constraints of three kinds. First, user identity, authority, and access permissions. Second, who a piece of information is attributed to, and how long it remains valid. Third, the rules for resolving requirements that conflict between users, and what it takes to complete a joint decision. These constraints shape what information or decisions must be obtained before an action can proceed, and what conditions must hold while it executes.

Org-Agent is built around that observation. It is a unified, constraint-centric reasoning framework that organises task execution into three stages, so that the constraints are honoured by the structure of the work rather than left to chance.

In the first stage, Org-Agent decomposes a task into atomic subtasks and constructs a task dependency graph, whose edges encode the dependencies among them. In the second, it schedules the subtasks in dependency order through topological sorting, so nothing runs before the things it depends on. In the third, it executes each subtask while accounting for the task's constraints, supported by tools for acquiring evidence and for managing memory. Planning the order comes first, and only then does anything get done. The graph captures what must happen before what, and the execution stage is where the constraints themselves are checked.

Experiments on MUSES-Bench and GroupMemBench show the framework is effective on both capabilities. Ablations further support what dependency modeling and tool use each contribute.

Going from one user to many brings authority, attribution, and conflict into play. A personal assistant can take one person's word as final. An organisational agent has to ask whose word it is, whether it is still current, and whether someone else's request overrides it. Org-Agent's response is to make the order of work explicit, and to check the constraints at every step along that order.