Org-Agent: constraint-centric task graphs for agents that serve whole organizations, not one user

Org-Agent: Beyond Personal Assistants Towards Organizational Agents

Luyao Zhuang, Yujing Zhang, Zijin Hong, Yilin Xiao, Xiao Huang

cs.AI

2026-09-28

Org-Agent splits organizational tasks into atomic subtasks ordered on a dependency graph under organizational constraints, validated on MUSES-Bench and GroupMemBench.

What problem this solves

Most agent frameworks assume a single user: one person, one context window, one memory that belongs to that person. Inside an organization those assumptions break. A request from user A may depend on information user B provided last week; some documents are visible only to certain roles; two people can issue conflicting requirements for the same task. Frameworks built for the personal-assistant setting do not model any of this, so the agent either misses information, oversteps its authority, or improvises when requirements collide.

The paper names two capabilities this setting demands: cross-user interaction and decision-making (turning several people's requirements into one executable, compliant decision), and cross-user memory and knowledge use (using information scattered across separate conversations at the right time, with the right permissions).

Method

Org-Agent takes a constraint-first stance: determine what information or approvals must be secured before an action and what conditions must hold during it, then plan execution. Constraints fall into three families: user identity, authority, and access permissions; attribution and temporal validity of information (who said it, when, and whether it still holds); and rules for resolving conflicting requirements plus completion requirements for joint decisions.

Execution runs in three stages:

• Decompose the task into atomic subtasks, each small enough to execute and verify on its own.
• Build a task dependency graph whose edges encode dependencies, then schedule subtasks in dependency order via topological sorting.
• Execute each subtask against the constraints, supported by two tools: evidence acquisition, which recovers the provenance of information, and memory management, which reads and writes shared cross-user memory.

The dependency graph does real work here. Organizational constraints inherently impose order: confirm with the person holding authority before reading the file, resolve a conflict before proceeding. Encoding constraints as graph edges gives the schedule a basis instead of leaving ordering to the model's improvisation.

Results

The paper evaluates on two benchmarks, MUSES-Bench and GroupMemBench, reporting that Org-Agent is effective on both capabilities, with ablations supporting the contributions of dependency modeling and tool use. The abstract contains no scores and names no baselines, and the full text was not retrieved for this summary, so 'by how much, against whom' cannot be answered here.

Why it matters

Permissions, audit trails, and provenance are compliance requirements in any real deployment, and most agent stacks treat them as an afterthought bolted on at the serving layer. Org-Agent moves them inside reasoning and scheduling, which is the right direction as agents go from personal tools to organizational infrastructure. Even without adopting the framework, the principle that constraints determine execution order is reusable. Said plainly: from the abstract this is a systems contribution combining mature parts (decomposition, graph scheduling, tool calls); the novelty lies in the combination and in the systematic definition of organizational constraints.

Limitations

• Only the abstract was available. Scores, baseline setups, benchmark composition, and ablation magnitudes all sit in the unretrieved full text, so the word 'effective' currently has no numbers behind it.
• Where do constraints come from? The abstract does not say whether they are supplied explicitly with each task or must be inferred from organizational rules and past interactions. Real organizational constraints are mostly implicit, and inference is the hard part.
• Topological sorting assumes dependencies are known when the graph is built. Dependencies discovered mid-execution, such as learning only after acquiring evidence that someone else must be asked, are not addressed.
• Whether MUSES-Bench and GroupMemBench were introduced by this same paper is not clear from the abstract. If benchmark and method share authors, the results deserve a discount.

Terms

• Organizational constraints: the paper's umbrella for identity/authority/permissions, information attribution and temporal validity, and conflict-resolution rules.
• Atomic subtask: a task decomposed to a single step that can be executed and verified independently.
• Task dependency graph: a directed graph whose nodes are subtasks and whose edges encode ordering dependencies.
• Topological sorting: an algorithm that linearly orders a directed graph so every predecessor precedes its successors.

Source

• arXiv 2609.34392v1
• PDF
• Hugging Face

Related papers

• Naive RAG Beats 12 Agentic Memory Systems, Ranking First in 6 of 7 Long-Memory Benchmark Settings
• Five seller models inflate tool fields; Gemini 3.8 Flash cuts success from 89% to 30%
• 273 idea atoms distilled from LLM papers drive a sampler toward directions no community would propose
• Meta Experts Match Vanilla MoE Loss at 70-80% Expert Budget
• Throwing away synthesized agent rules beats stored abstractions, 59.9% vs 49.0%
• Jev 1.13 lifts terminal-agent cost 19% on Terminal-Bench; pass rate stays 73%

All paper explainers