{"id":"d8fe5e48-346c-4cd8-9a72-599fd1809131","arxiv_id":"2608.10915","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A new human-centric agentic AI paradigm, Combodied Agents, organizes perception, memory, prediction, and intervention around the evolving human state and agency over time.","lead":"This paper proposes a new class of AI agents, Combodied Agents, that track and support a person's state over time using sensors, software, robots, and human services as tools. It argues that agentic AI should be judged by long-term human benefit and preserved agency, not just task completion, and lays out a framework for building and testing such agents.","discovery_kind":"paradigm_shift","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (8) lets the intervention policy maximize expected utility under a predictive, non-causal PWM, while Section 4.2 admits that observed user behavior is confounded; the closed loop can therefore select interventions on spurious correlations unless Eq.","rationale":"I read the paper as a paradigm proposal, not an empirical systems paper. Its central claim is that organizing agentic AI around longitudinal human-state trajectories will yield safe, beneficial support. For that claim to hold, the formal closed loop must at least be consistent. The reader's weakest assumption concerned empirical feasibility of PWMs from sparse data; that is real and acknowledged in Sections 4.3 and 8.4. But there is a sharper problem in the formalization itself. Eq. (8) is the paper's canonical statement of how an intervention is chosen, and it does not distinguish predictive from causal PWMs. The paper itself draws this distinction in Section 4.2 and warns about confounding, yet the selection rule does not enforce the distinction. This is not a lack of experimental validation; it is a gap in the definition of the paradigm. If a predictive PWM is allowed to drive interventions, then even a perfectly learned, well-calibrated predictive model can systematically choose actions that look beneficial but are merely correlated with good outcomes. The paper's safety constraints—consent, reversibility, uncertainty thresholds—are valuable but they do not repair a biased objective. A revised framework should either restrict Eq. (8) to causally identified distributions or explicitly separate a 'decision-support' mode from an 'autonomous intervention' mode. The concrete test above would settle whether the current formalization is merely underspecified or actually selects harmful interventions under confounding. I therefore concur with the reader's CONDITIONAL verdict: the paper should be accepted only with the formal amendment and an empirical demonstration of intervention-effect identifiability. Since the reader already recommends CONDITIONAL, my stress-test does not change the verdict, but it tightens the condition.","tokens_in":36900,"tokens_out":6990,"duration_ms":62385,"concrete_test":"Run the Section 2.4/4.2 loop on a micro-randomized trial dataset (e.g., HeartSteps) as follows: (1) train a predictive PWM p^a_theta on the observational pre-randomization history only; (2) train a causal PWM using the randomized intervention assignments as instruments; (3) compute Eq. (8) with each model and compare the selected interventions against the known true causal effects from the trial. If the predictive-PWM policy selects any action whose doubly robust causal effect estimate is negative or non-significant while the causal-PWM policy does not, the framework as formalized is unsafe without requiring causal identifiability. The paper should then amend Eq. (8) to require p^a_theta to be a causally identified interventional distribution for any action in A_adm that changes the user's state.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The canonical selection rule in Section 4.2, Eq. (8), computes a* in ParetoArgmax over A_adm of E_{p^a_theta}[U(Y,G_t)], where p^a_theta is the PWM's scenario-conditioned distribution. The paper correctly notes in Section 4.2 that 'a predictive PWM compares scenario distributions under explicit assumptions' and that a 'stronger causal PWM' is needed for interventional queries, and it warns that observed behavior is 'confounded by motivation, hidden context, health status, prior interactions, and selective engagement.' However, Eq. (8) does not restrict p^a_theta to a causal, identified model. As written, the admissible policy can select an intervention based on a conditional association: for example, if reminders are preferentially issued to already-adherent users, a predictive PWM will associate reminders with better adherence and Eq. (8) will keep selecting them even though the causal effect may be nil or negative. The A_adm constraints (consent, uncertainty, reversibility) reduce harm but do not correct a biased expected-utility ranking. The central claim that Combodied Agents 'support' human trajectories therefore depends on an unstated formal requirement—that p^a_theta encodes intervention effects, not merely conditional predictions—which is nowhere enforced in the framework and is deferred to the open problem of 'causal intervention learning' in Section 8.4. This is a formal under-specification of the closed loop, not just an empirical feasibility gap.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript introduces Combodied Agents as a human-centric paradigm of Agentic AI, defined as agents that perceive, model, and influence the evolving state of a person through continuous multimodal sensing and longitudinal interaction. The paper develops a closed-loop formalization in Section 2.4, in which multimodal event evidence feeds a latent human-state posterior, a longitudinal memory, a Personal World Model (PWM) that predicts state-event-outcome trajectories under candidate interventions, and an admissible intervention policy constrained by consent, uncertainty, safety, reversibility, and user control. It then surveys sensing modalities and data-quality issues, positions PWMs relative to profiles, memory, and generative agents, proposes edge-native deployment stages, and lays out scenario-centered evaluation, agency-preservation metrics, a taxonomy of human-state targets, relationship modes, and agent roles, and a research agenda of open problems. The paper is explicitly a position and formalization paper: it reports no experiments, and Sections 6 and 8.4 state that empirical validation and methods for learning individual dynamics and causal intervention effects remain future work.","tokens_in":37220,"tokens_out":4692,"duration_ms":50361,"significance":"If the framework is taken as a design agenda rather than an empirical result, it makes a useful contribution by consolidating fragmented capabilities in personal assistants, health agents, companions, and adaptive human-AI systems under a single longitudinal, human-state-centered objective. The explicit distinction between observations, reconstructed events, inferred states, predicted trajectories, and authorized interventions is a valuable formal discipline, and the agency-preservation metrics in Section 6.3 go beyond typical engagement-based evaluation. The paper is also honest about its own limits: Sections 4.3 and 8.4 explicitly defer the load-bearing learning and causal-inference problems. The contribution is therefore primarily conceptual and organizational; the central empirical claim that Combodied Agents improve long-term human outcomes is not yet supported by data and, as discussed below, the formal selection rule in Eq. (8) is under-specified with respect to causal validity.","major_comments":[{"comment":"The stress-test concern is valid and lands. Eq. (8) selects a* in ParetoArgmax over A_adm of E_{p^a_theta}[U(Y,G_t)], but Eq. (7) defines p_theta as a predictive distribution over future states and outcomes under a candidate scenario, not as an interventional distribution. The paper itself acknowledges in the discussion of Eq. (9) that observed user behavior is confounded by motivation, hidden context, health status, prior interactions, and selective engagement, and that writing do(·) does not remove confounders. As written, Eq. (8) can therefore choose an intervention because the PWM captures a conditional association rather than a causal effect; for example, if reminders are issued preferentially to already-adherent users, the policy will infer that reminders improve adherence when the causal effect may be nil or negative. The admissibility set A_adm constrains consent, uncertainty, and reversibility but does not correct a biased expected-utility ranking. This is a formal under-specification of the closed loop, not only an empirical feasibility gap. The framework should either require that the expectation in Eq. (8) be taken under an identified causal/interventional distribution, or restrict the admissible set to actions whose effects are supported by a causal model with explicit identification assumptions; otherwise the central claim that a Combodied Agent 'supports' the human trajectory is not entailed by the framework.","section":"Section 4.2, Eq. (8)"},{"comment":"The feasibility of learning person-specific dynamics and intervention effects is a load-bearing premise of the paradigm. Section 4.3 states that \"Sparse personal data make training a model from scratch inappropriate for most users,\" and Section 8.4 lists \"learning individual dynamics\" and \"causal intervention learning\" as open problems, while Section 4.2 notes that high-risk systems must not conduct unconstrained exploration. The paper does not provide a concrete argument or small-scale demonstration that a PWM can be learned from realistic sparse, noisy, multimodal data with calibrated uncertainty and safe estimation of intervention effects. Since Section 6 proposes CombodiedBench but does not instantiate it with a single worked scenario or baseline, the empirical core of the proposed closed loop remains promissory. I am not asking for a full deployment study in a position paper, but the manuscript should either include a worked synthetic or retrospective case study, or give a precise feasibility argument (for example, what population prior plus limited personal adaptation yields what bound on personalization error, and under what identification assumptions). Without this, the claim that the loop can deliver \"proportionate support\" is not yet supported.","section":"Section 4.3 and Section 8.4"},{"comment":"The benchmark proposal is described only at the level of module names and general scoring principles; no concrete scenario episodes, metric definitions, action spaces, or evaluation protocols are specified. This matters because Section 6.2 and Table 8 claim that evaluation should connect longitudinal person modeling, intervention effect, and human outcome, but the paper provides no instantiation that would let a reader reproduce a single evaluation. Combined with the absence of any experimental results, the abstract's assertion that Combodied Agents \"aim to improve health, learning, judgment, capability, relationships, and goal pursuit\" is an untested programmatic statement rather than a substantiated claim. I recommend either adding one fully worked benchmark episode with explicit scoring and baseline comparisons, or explicitly and consistently framing the paper's contribution as a formal and design framework whose empirical validation is entirely future work; the latter framing is currently implied in several places but not maintained throughout the abstract and introduction.","section":"Section 6.4, CombodiedBench"}],"minor_comments":[{"comment":"The notation E_{p^a_theta} in Eq. (8) is used before p^a_theta is defined; Eq. (7) defines p_theta. Please define the superscript-a notation explicitly so that the expectation is unambiguous.","section":"Section 4.2, Eqs. (7)-(8)"},{"comment":"The acronym \"PPL\" appears in the training-objective line of Figure 3 but is never defined; if it means perplexity, say so in the caption or text.","section":"Figure 3"},{"comment":"The figure contains the typo \"Labtop\" where \"Laptop\" is intended; please correct it.","section":"Figure 4"},{"comment":"The conceptual distinction between delegation-based and growth-based agency relies heavily on an author's GitHub repository [12] and a July 2026 blog post [13]; these are not peer-reviewed sources, and the paper would benefit from additional peer-reviewed references on human-centered AI, dependence, and capability development.","section":"References [12] and [13]"},{"comment":"The chain \"observation to event to inferred state to predicted trajectory to authorized intervention\" is useful and should be referenced more prominently in the formal sections, since it directly supports the later causal-validity concern.","section":"Section 3.10"}],"recommendation":"major_revision","confidential_remarks":"This is an agenda-setting and formalization paper rather than an empirical contribution. The core idea is coherent and timely, but the load-bearing formal issue in Eq. (8) needs to be addressed before the framework can support its central claim, and the empirical feasibility argument is currently deferred. The paper is also very long and would benefit from tighter focus on the formal loop and one fully worked evaluation scenario. If the journal does not usually publish position papers without some proof-of-concept, the fit should be considered carefully. The authors' reliance on their own unpublished framework [12] for a foundational conceptual distinction may also draw scrutiny from reviewers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this paper if you care about where agentic AI is heading. It makes a genuine move: shifting the primary action substrate from digital or physical state to the evolving human state, and then building a closed loop around that substrate. None of the components is new, but the integration is, and the paper executes it with unusual care. The taxonomy (human-state target, relationship mode, agent role), the evaluation matrix with non-compensatory failures, and the edge-native deployment discussion are all useful. The authors are also honest: Section 6 says no empirical validation is claimed, and Section 8.4 lists learning individual dynamics and causal intervention learning as open problems.\n\nThe soft spots are the ones the authors themselves point at. The load-bearing premise is that a Personal World Model can learn person-specific dynamics from sparse, noisy data with calibrated uncertainty; that is not demonstrated anywhere, and the paper knows it. The stress-test note about Eq. (8) is fair: as written, the policy maximizes expected utility under a predictive, possibly confounded p_a_theta, so it could pick interventions based on spurious correlations even though the text warns about confounding. The authors explicitly say a stronger causal PWM is needed for interventional queries, but they do not enforce that in the canonical selection rule. That is a formal gap, not just an empirical one. It does not sink the paradigm; it means the framework should be read as a proposal with an explicit boundary. A referee should ask them to either restrict Eq. (8) to identified causal models or state more clearly that the policy is only valid for descriptive scenario comparison until such models exist.\n\nMinor complaint: reference [12] is a GitHub repository rather than a peer-reviewed source, and several 2026 arXiv references will age quickly. The paper is long but well organized.\n\nWho should read it: anyone designing human-centered agents or benchmarks for them. It does not deserve a desk reject; it deserves a serious referee, with the expectation that the formalization of causal selection gets tightened. I would accept it for review. Reading group: maybe, if you have a session on agent paradigms. I would cite the taxonomy if writing about evaluation.","headline":"A careful, honest paradigm proposal for human-centered agentic AI, with a real formal gap around causal selection that the authors themselves flag.","tokens_in":37776,"tokens_out":2754,"would_cite":true,"duration_ms":27374,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper introduces Combodied Agents, a human-centered paradigm of agentic AI that makes a person's evolving state and agency—not an external task—the primary object of perception, prediction, intervention, and evaluation.","keywords":["Combodied Agents","human-centric AI","agentic AI","Personal World Model","agency preservation","longitudinal memory","intervention policy","edge-native personal models"],"falsifier":"Run a pre-registered longitudinal study of a medication-adherence or mood-support agent in which the PWM's scenario-conditioned predictions (adherence, symptom state, or wellbeing under intervention policy A versus policy B) are recorded before outcomes are observed; the paradigm's empirical core is refuted if those predictions do not beat a population-baseline model on calibration and scenario discrimination, or if the two policies' predicted trajectories are indistinguishable.","tokens_in":36747,"feed_emoji":"👤","tokens_out":13958,"duration_ms":104324,"temperature":0.7,"pith_summary":"This paper introduces Combodied Agents, a proposed third center of gravity for agentic AI alongside digital agents (which transform software states) and embodied agents (which transform physical states). A Combodied Agent—the name joins \"companion\" and \"body\"—makes the evolving state of a particular person, including health, cognition, emotion, behavior, relationships, and goals, the primary object of modeling, intervention, and evaluation. The authors formalize a closed loop in which event-based multimodal perception feeds a longitudinal, correctable memory; a Personal World Model predicts how the person's state would unfold under alternative decisions and interventions; and an admissible intervention policy chooses proportionate support under consent, uncertainty, safety, reversibility, and user control. If the paradigm holds, agentic AI would be judged not by how many tasks it takes off the human's hands, but by whether repeated use leaves people more capable, autonomous, and healthy over time.","feed_headline":"Shift agentic AI: model the person's trajectory, not just tasks","feed_subtitle":"It joins perception, memory, personal world models, and safe intervention to keep people more capable over time.","key_machinery":"The load-bearing object is the Personal World Model (PWM), defined as a purpose-bounded, individual-specific event-dynamics model that turns a person's governed history of multimodal event-evidence records into calibrated distributions over future human states, observable events, and scenario-relevant outcomes under alternative decisions and interventions. Around it runs a formalized closed loop: a posterior state representation $Z_t$ is inferred from event evidence $D_{\\leq t}$, longitudinal memory $M_t$, and context $C_t$; the PWM evaluates candidate scenarios $s_{t:t+\\Delta}$; and the policy selects $a^{\\mathrm{agent},*}_t \\in \\mathrm{ParetoArgmax}_{a \\in A^{\\mathrm{adm}}_t}$ over an admissible set that enforces consent, scope, safety, uncertainty, reversibility, and escalation, so that non-intervention, clarification, and referral are always available. The next human state $H_{t+1}$ depends on the agent's action, the user's own actions, and exogenous influences, and feedback updates memory and policy. This machinery is what upgrades personalization from retrieval and adaptation to intervention-conditioned prediction, and it is deliberately bounded: the paper rejects an exhaustive Human Digital Twin in favor of uncertainty-aware, user-correctable, purpose-limited representations.","core_discovery":"The paper's central claim is that the defining object of an agentic system should be the trajectory of the human subject, not the external task state. It defines Combodied Agents as human-centered intelligent agents that perceive, model, and influence the evolving state of a person through continuous multimodal sensing and longitudinal interaction, and it argues that no existing agent category integrates perception, memory, prediction, and intervention around that object. The technical core is the Personal World Model (PWM), a purpose-bounded, individual-specific event-dynamics model: given a governed history of event evidence, current context, and a candidate scenario, it outputs a calibrated distribution over future human states, observable events, and outcomes under alternative user actions and agent interventions. An intervention policy then selects among admissible actions, including non-intervention, clarification, and escalation, with agency preservation as a binding constraint. On this basis the paper reorganizes the design space by human-state target, relational context, and agent role, and proposes evaluation in which autonomy, contestability, capability growth, over-reliance, and relationship safety are reported along with task outcomes.","pith_inferences":["The framework gives a measurable operational meaning to otherwise vague debates about AI dependence and deskilling: an intervention's value is scored by its effect on the person's predicted future state-trajectory, so \"leaves you more capable\" becomes a benchmarkable quantity rather than a slogan.","The three-stage cloud-to-edge migration implies a testable architectural prediction: user correction, memory reversibility, and model portability will be the binding constraints on real adoption, because cloud-centric systems structurally cannot offer inspectable, migratable personal models.","Dependency and over-reliance—often treated as qualitative companion-safety concerns—could be operationalized as intervention-conditioned shifts in predicted trajectories, for example declining independent task attempts as the agent's availability is experimentally varied.","The paper's opening example, medication adherence in older adults, is also its most practical near-term test bed: event reconstruction, intervention response, and agency preservation can be measured there in a bounded, high-stakes setting within weeks rather than years."],"forward_implications":["Evaluation shifts from task-completion rates to longitudinal, non-compensatory metrics: capability growth, autonomy preservation, contestability, over-reliance risk, and relationship safety are reported alongside task outcomes.","Deployment moves toward edge-native personal models, with longitudinal memory, the PWM, and intervention authority resident on user-controlled devices and cloud services invoked selectively through a privacy gateway.","Existing agent categories—assistants, health agents, companions, learning tutors—are reclassified by human-state target, relational context, and agent role, with safety boundaries set by the relationship rather than by the interface.","Every intervention is gated by an admissibility constraint: consent, scope, safety, uncertainty, reversibility, and escalation bounds, with non-intervention, clarification, and referral as always-available actions.","A concrete benchmark program (CombodiedBench) becomes necessary, spanning perception, memory continuity, goal negotiation, intervention appropriateness, agency preservation, relationship boundaries, escalation, and longitudinal outcomes."],"supporting_citations":[{"why":"supplies the predictive world-model concept (an internal representation of how a state evolves under actions) that the Personal World Model redefines for a single person.","marker":"[84]"},{"why":"establishes the paper's motivating principle that what users retain and develop—understanding, competence, agency—is an outcome alongside task completion.","marker":"[12]"},{"why":"provides the long-term memory-augmented agent mechanism that the longitudinal memory component builds on and extends with correction and provenance.","marker":"[58]"},{"why":"supplies the long-term interactive memory benchmark that the evaluation section identifies as a relevant but partial resource.","marker":"[62]"},{"why":"demonstrates LLM reasoning over wearable and personal health data, the line of work the PWM is explicitly distinguished from as a reasoning interface rather than a learned dynamics model.","marker":"[42]"},{"why":"provides empirical evidence on companion-chatbot usage and loneliness that motivates the agency-preservation and dependency constraints.","marker":"[40]"},{"why":"offers the adjacent longitudinal health-agent framework whose adaptation, coherence, and agency emphases the paper synthesizes into its closed loop.","marker":"[87]"},{"why":"serves as the contrast case for the PWM definition, since generative agents simulate plausible behavior but lack the intervention-conditioned calibration contract.","marker":"[33]"}],"fun_headline_variants":["Personal world models put the person at agentic AI's core","From task completion to modeling human trajectories","Combodied agents: track the person, not just the task","Agentic AI that models your evolving state, not just tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework collapses if a Personal World Model cannot learn, from sparse, noisy, mostly observational data about a single person, how that person's states and outcomes respond to alternative interventions with trustworthy uncertainty estimates—without running unsafe experiments to obtain the missing data.","fun_headline_variants_meta":{"raw":{"variants":["Personal world models put the person at agentic AI's core","From task completion to modeling human trajectories","Combodied agents: track the person, not just the task","Agentic AI that models your evolving state, not just tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00076,"raw_usage":{"total_tokens":3429,"prompt_tokens":1052,"completion_tokens":2377,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":668,"completion_tokens_details":{"reasoning_tokens":2311}},"tokens_in":668,"tokens_out":2377,"duration_ms":16107,"temperature":1.0,"reasoning_tokens":2311,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:21:13.935202+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a pre-registered longitudinal study of a medication-adherence or mood-support agent in which the PWM's scenario-conditioned predictions (adherence, symptom state, or wellbeing under intervention policy A versus policy B) are recorded before outcomes are observed; the paradigm's empirical core is refuted if those predictions do not beat a population-baseline model on calibration and scenario discrimination, or if the two policies' predicted trajectories are indistinguishable.","supporting_citations":[{"cited_title":"A framework for longitudinal health ai agents.Nature health, pages 1–10, 2026","cited_arxiv_id":null,"evidence_quote":"offers the adjacent longitudinal health-agent framework whose adaptation, coherence, and agency emphases the paper synthesizes into its closed loop."}],"review_version":1}