Review Comment:
This work introduces the concept of Computer-Using Personal Agents (CUPAs), which extend standard Computer-Using Agents (CUAs) by integrating them with Personal Knowledge Graphs (PKGs). The authors propose that PKGs can serve as structured, semantic, and user-controlled repositories for private data, allowing agents to automate complex, privacy-sensitive tasks while adhering to user-defined access policies. To motivate this, the paper presents two relevant primary user scenarios: planning a dinner date and personalising a lifelong learning coach; and subsequently outlines the state of the art, necessary agent capabilities, technical challenges, and a roadmap for future research.
Overall, the vision articulated in the paper is clear, and the authors nicely highlight the growing need for user-centric, privacy-preserving personal agents. The contribution is highly relevant and aligns very closely with the existing vision and established trajectory of PKGs. In this sense, while the core conceptualisation is not entirely novel, it is certainly valuable to see the authors elaborating on this paradigm in parallel with the rapidly evolving landscape of generative AI and agentic systems. The writing is generally of good quality, but I found the execution of the narrative and the depth of the technical discussion to occasionally limit the paper's overall significance and impact.
Major Comments
The primary weakness of the manuscript lies in its structural choices and the fact that technical arguments appear occasionally as high level.
(*) Firstly, the structure is often confusing and potentially repetitive, particularly regarding the multiple sections named "User Scenarios" used to contextualise the narrative. Currently, the initial scenarios immediately present a CUPA-based solution, intertwining the problem with the proposed resolution. To strengthen the contribution, it would be highly beneficial to disentangle the scenario definitions from the solutions. I suggest first presenting the scenarios purely as use cases detailing the specific needs of each user/application without disclosing CUPA. Then, after introducing CUPA's main characteristics, the authors could contrast an implementation using today's GenAI technologies (e.g., agentic methods, LLMs) against a CUPA-based solution to explicitly highlight the differences and give value to each approach.
(*) Secondly, the remarks in the "CUPA Capabilities" section are often superficial and require substantiation. Arguments such as "the agent should avoid acting in an unethical or illegal manner" apply to any agent, raising unaddressed questions about how a CUPA determines legality, manages relevant versus irrelevant pattern learning, or handles derivative data ownership from a PKG. Rather than attempting to redefine design goals and risking an incomplete or superficial set of requirements, the authors should ground CUPA within an established framework for Trustworthy/Responsible AI (such as the AI HLEG Guidelines) to systematically address these technical challenges.
Minor Comments
(*) The mention of the Solid Pod on page 3 requires a supporting reference or footnote (a reference is provided later).
(*) Listing 1 is very useful for supporting the narrative, but it is sometimes hard to read. Although it is a simple example, the fragment contains a few inconsistencies in its use (e.g., the declaration of the study plan) and editing (e.g., comment line breaks).
(*) Figure 2 appears slightly cluttered, and the intended reading flow or sequence is not immediately clear.
(*) In the "Data Sovereignty & Governance" section (page 5), various SSI systems and personal AI agents like kwaai.ai are introduced, mentioning that CUPAs also strive to be a Sovereign Digital Architecture (SDA). The authors should clarify if this is a shared characteristic between CUPAs and these other systems. Additionally, the quoted text regarding Samsung’s acquisition of Oxford Semantic Technologies lacks a source and must be properly cited. The suggested revision of the narrative should facilitate this comparison.
(*) Some expressions throughout the manuscript are overly informal, such as the use of "nasty surprises" (page 6, column 1).
(*) The heterogeneity and multimodality challenge presented on page 8 could be better linked to current work in multimodal foundation models. The ability to acquire and process multimodal streams of information could be positioned as part of the solution rather than a competing factor. Furthermore, the paragraph discussing the limitations of "NLP methods" feels somewhat speculative and outdated, as the field has moved beyond traditional NLP, and it is missing relevant contemporary citations.
(*) On page 8, column 2, the text states that "A CUPA agent manages the creation, maintenance, and validation of a PKG," whereas earlier sections imply that the user maintains the PKG using a data store like Solid, with the CUPA merely accessing and reusing (or enriching) specific parts. This discrepancy should be reconciled.
(*) Regarding "Web-aware CUPAs", it would be useful to clarify whether LLM agents equipped with Web search features (i.e., those parsing HTML content directly) would also fit within this category.
(*) Under "Ethical and Societal Concerns", since CUPAs are AI software systems, the authors could ground them in the current regulatory landscape (e.g., the EU AI Act or similar legal frameworks) to address the anticipated ethical and societal challenges.
| 
