Combodied Agents: A New Paradigm of Human-Centric Agentic AI

Qianggang Ding, Xingyao Wang, Rui Feng, Zhibin Wang, Feixiang Yao, Kelong Mao, Hao Sun, Zhiyao Luo, Jiankai Tang, Lei Li, Jiadong Guo, Minheng Ni, Weicong Lin, Chenxi Yang, Hongxiang Gao, Zhenghua Chen, Yang Bai, Min Wu, Jun Cheng, Huazhu Fu, Dacheng Tao, Bang Liu

Combodied Agents shift AI from task-completion to supporting human-state trajectories over time.

What defines "Combodied Agents"—systems that act on the human subject over time—and how do they differ from existing digital or embodied AI?

Current AI agents are designed to either manipulate digital interfaces or control physical robots, but they lack a framework for understanding the evolving state of the human they serve. This leads to systems that optimize for task completion while ignoring long-term impacts on human agency, health, and capability. Combodied Agents reorganize AI around the human subject: they integrate multimodal perception, longitudinal memory, and personal world models to provide adaptive, consent-aware support. The core move is to treat the human’s physiological, cognitive, and behavioral trajectory as the primary target of intervention rather than an external software or physical state. This paradigm aims to improve human outcomes—such as health, judgment, and goal pursuit—without treating maximum automation or user dependence as measures of success.

Paper Primer

The authors identify a structural gap in Agentic AI: Digital Agents focus on software states, and Embodied Agents focus on physical states, but neither treats the human as the primary object of modeling. This creates a "substitution trap" where AI completes tasks while weakening the user's own ability to judge, decide, or develop expertise.

Combodied Agents function as a closed-loop system: they reconstruct personal events from multimodal data, store them in a correctable longitudinal memory, and use a Policy-World-Model to predict outcomes under various interventions. The system acts like a navigator for a person's life: it reads the current "terrain" (human state), checks the "map" (longitudinal memory), and suggests a path that preserves the user's autonomy rather than taking the wheel entirely.

Combodied Agents improve long-term human capability by shifting the optimization target from isolated task success to longitudinal human gain.

The framework replaces task-centric metrics with agency-preservation metrics, evaluating whether interventions support understanding, competence, and calibrated reliance. The paradigm redefines success as sustained human benefit, explicitly rejecting maximum automation as a proxy for agent capability.

Why is this a new paradigm rather than just a more advanced personal assistant?

Existing assistants are fragmented across memory, sensing, and domain-specific support. Combodied Agents unify these into a closed loop that models the human's evolving state trajectory, ensuring that interventions are proportionate, reversible, and constrained by the user's consent and long-term wellbeing.

Does this require a complete digital replica of a person?

No. Unlike a Human Digital Twin, which aims for high-fidelity simulation, Combodied Agents use purpose-bounded, uncertainty-aware, and user-correctable representations of the specific aspects needed for an agreed support context.

The Case for Combodied Agents

We define Combodied Agents, a human‑centric AI paradigm that models and supports evolving human states.

Current agentic AI either manipulates digital states (Digital Agents) or physical states (Embodied Agents), but neither treats the human’s evolving state as the primary target. This leaves a structural gap: after a missed medication dose, an agent can remind or fetch the pill without understanding why the person failed to take it. The paper therefore proposes a new paradigm that centers on modeling and supporting longitudinal human‑state trajectories.

A Combodied Agent is an AI system whose chief objective is to improve a person’s future state over time, rather than merely completing a predefined task.

The closed‑loop architecture consists of event‑based multimodal perception, longitudinal memory, a Personal World Model that turns evidence into calibrated future‑state distributions, and an admissible intervention policy that respects consent and safety. Feedback from the person and environment continuously updates each component, enabling proportionate support without requiring an exhaustive Human Digital Twin. This purpose‑bounded, uncertainty‑aware, user‑correctable representation makes the system tractable while still delivering sustained human benefit.

**Table 1.** Consolidated related-agent landscape: representative capabilities, limitations, and the human-centered longitudinal support added by Combodied Agents.

Shifting from task‑centric to human‑centric agentic AI reorients success metrics toward sustained human benefit and agency preservation.

Defining the Combodied Paradigm

Defines Combodied Agents, distinguishes them from Digital and Embodied agents, and maps their core capabilities.

Across the surveyed landscape, agents are clustered by the state they aim to transform: digital artifacts, physical bodies, or the evolving human person.

Digital Agents are systems whose primary action substrate is the manipulation of digital states—files, code, UI elements, or API calls. They operate entirely within software environments, using perception of screen content, logs, or structured data to infer task progress and then issue actions such as clicks, commands, or data transformations. Evaluation typically measures correctness, completion rate, latency, and policy compliance.

Embodied Agents interact with the physical or simulated physical world. Their action substrate is a body—robotic limbs, vehicle actuators, or simulated avatars—that directly changes spatial configurations, forces, or material states. Perception comes from vision, lidar, proprioception, or simulation feedback, and the internal model focuses on geometry, dynamics, and safety constraints. Success is measured by task completion, safety incidents, and generalization across bodies and environments.

Combodied Agents close the loop on the evolving human person. Their action substrate is the human’s physiological, cognitive, emotional, and social state. They fuse multimodal sensing, longitudinal memory, and a Personal World Model (PWM) to predict how interventions will affect future human outcomes. Interventions span information, reminders, coaching, protection, and escalation, always respecting consent and agency preservation. Evaluation targets well‑being, autonomy, and long‑term health or capability gains.

Combodied Agents operate in a closed‑loop: observations feed belief updates, which inform goals, plans, and actions that affect the human state, producing new observations. This loop is formalized by the sequence oₜ → bₜ → (gₜ, pₜ) → aₜ → xₜ₊₁ → oₜ₊₁ and extended with latent state variables $H$ₜ, memory Mₜ, and the policy‑world‑model (PWM).

A Digital Agent focuses on changing digital artifacts or information states, acting through software interfaces without direct physical effect.

How does a Digital Agent differ from a conventional software assistant?

A conventional assistant may simply retrieve information or generate text, whereas a Digital Agent actively modifies digital state (e.g., edits code, triggers workflows) and maintains a closed‑loop belief about task progress.

An Embodied Agent’s primary effect is on physical or simulated bodies, using perception‑action loops that respect dynamics and safety.

Why isn’t an autonomous vehicle simply a Digital Agent?

Because its target state is the vehicle’s physical pose and surrounding environment; it must reason about dynamics, collisions, and kinematics—concerns absent in purely digital state manipulation.

**Table 1.** Consolidated related-agent landscape: representative capabilities, limitations, and the human-centered longitudinal support added by Combodied Agents.

The table compares "Assistive Care Agents" and "Edge AI Agents" across three columns: "Capabilities", "Limitations", and "What Combodied Agents Add".

The image illustrates a conceptual framework for human-centric state modeling. At the center is a human silhouette representing the "Human State," surrounded by icons for "Cognition," "Behavior," "Context," "Emotion," "Body," and "Human Agency." A flow of information moves from the "Past" through a "Perceive" stage, into the central human model, and then through an "Act With" stage toward "Futures." Below the diagram, five key properties are defined: Human-centric state modeling, Longitudinality, Intervention, Co-agency, and Agency preservation.

**Figure 1.** Action substrates and representative task configurations in Agentic AI. (a) Representative substrate-dominant and cross-substrate tasks for Digital, Embodied, and Combodied Agents. The examples are illustrative rather than exhaustive, and their placement indicates which target states and evaluation objectives are jointly involved. (b) The three overlapping centers of gravity are defined by the states that primarily organize modeling, action, and evaluation: digital states and artifacts, physical or simulated physical states, and evolving human states and agency. Pairwise overlaps denote systems that couple digital planning with physical execution, digital support with longitudinal person modeling, or physical assistance with adaptation to human outcomes; the central overlap integrates all three. All three paradigms draw on shared agentic machinery for perception, representation, reasoning, planning, action, learning, and evaluation.

Multimodal Perception for Human States

Multimodal perception aggregates fragmented personal data into event‑evidence for longitudinal modeling.

Across the surveyed studies, multimodal perception is unanimously treated as an event‑centric pipeline: raw fragments are filtered, aligned, and interpreted before entering longitudinal memory.

Longitudinal Memory stores time‑stamped event‑evidence records so that a Combodied Agent can retrieve past observations, their interpretations, and associated uncertainties when reasoning about current or future states.

How does Longitudinal Memory differ from a conventional short‑term cache?

Unlike a cache that merely holds recent raw samples, Longitudinal Memory preserves the full provenance of each event—who/what generated it, under what acquisition mode, and with what uncertainty—so the agent can later audit, reinterpret, or delete the record according to user consent.

Textual inputs (messages, diaries, questionnaires, notes) provide explicit statements of goals, preferences, emotions, and consent. The survey reports that user‑reported and device‑mediated acquisition modes differ in evidential strength: self‑authored goals are high‑certainty, while imported notes carry lower provenance.

Audio captures both linguistic content and paralinguistic cues (pitch, pauses, ambient sounds). Acquisition can be near‑field (phone, earbuds) or far‑field (room microphones). The paper notes that continuous ambient listening raises privacy risk, while triggered capture yields higher‑quality event evidence.

Visual streams (RGB, depth, infrared, eye‑tracking) enable detection of activities, gestures, and environmental context. The survey distinguishes device‑facing, wearable egocentric, and fixed ambient cameras. Privacy‑preserving pipelines replace raw frames with event descriptors before storage.

Wearables and clinical devices deliver heart‑rate, HRV, SpO₂, EDA, respiration, temperature, glucose, and emerging biochemical markers. Acquisition mode (continuous wear vs. episodic patch) determines burden and signal fidelity.

Inertial sensors, GPS, BLE, Wi‑Fi, radar, and smart‑home devices capture activity, location, and routine patterns. The survey groups configurations by placement (wrist, phone, ambient) and sampling strategy (continuous vs. event‑triggered).

Metadata from calendars, communication logs, proximity sensors, and caregiver interactions encode relationships and coordination. Acquisition can be active (user‑reported) or passive (ambient proximity).

Ambient sensors and external services report light, noise, temperature, air quality, occupancy, weather, and connectivity. These modalities rarely constitute primary evidence but act as disambiguators for other signals.

Electronic health records, legal documents, educational transcripts, and employment data arrive via standards (FHIR, OMOP) or user uploads. They are episodic, high‑authority, and often delayed.

**Figure 2.** Functional organization of a Combodied Agent. Multimodal observations support human-state estimation; longitudinal memory supplies relevant temporal context; a personal world model predicts state–event–outcome trajectories under alternative decisions, interventions, and contextual changes; and an intervention policy selects and delivers appropriate support. Feedback from the person and environment closes the loop. Safety, uncertainty, consent, and user control constrain every stage.

Table 4 enumerates the seven memory components (episodic, semantic person, trajectory, goal, relationship, intervention‑response, user‑control) that together constitute Longitudinal Memory. Table 6 details acquisition configurations for each modality, highlighting how sensor choice, sampling mode, and event focus jointly shape coverage, burden, and privacy.

Event‑based multimodal fusion proceeds by filtering modality‑specific fragments, aligning them temporally, and reconstructing a coherent evidence chain. The process respects provenance at every step, allowing the agent to retain uncertainty and to defer action when contradictory or low‑confidence evidence persists.

The Personal World Model

Personal World Models unify prediction, intervention, and agency for individualized long‑term assistance.

Across personal‑modeling research, the field converges on a hierarchy: static profiles, memory stores, adaptive agents, and finally full Personal World Models that predict intervention‑conditioned trajectories.

A PWM is a purpose‑bounded, probabilistic model that predicts how an individual’s state evolves under specific interventions.

How does a PWM differ from a generic personal profile?

A profile only stores static attributes (age, preferences). A PWM actively predicts how those attributes will change under concrete actions, quantifies uncertainty, and supports counterfactual scenario comparison.

Intervention semantics encode actions as conditional variables that shape the PWM’s predicted future trajectory.

What does “intervention‑conditioned modeling” entail versus a simple action recommendation?

Instead of merely suggesting an action, the model treats the action as a conditioning variable, predicts its downstream impact on the person’s state, and quantifies the resulting outcome distribution.

Static representations that capture a person’s attributes, preferences, and demographic information. They are typically learned from enrollment data or occasional surveys and remain unchanged unless explicitly updated.

Memory modules that store a time‑ordered record of observed events, behaviors, and signals. They enable recall, provenance tracking, and correction of past observations.

Agent systems that adapt outputs, plans, or actions to an individual user based on the available personal data. They typically optimize a task‑specific utility (e.g., recommendation relevance).

Models that generate plausible future behaviors or dialogues for a person, often using large language models or simulation frameworks. Their primary goal is realism rather than accurate outcome prediction.

A purpose‑bounded, individual‑specific dynamics model that predicts how a person’s state‑event trajectory evolves under alternative decisions, interventions, and contexts, while explicitly representing uncertainty.

**Figure 3.** Reference technical scheme of a PWM. Event evidence and longitudinal memory support a posterior over latent personal state; domain- and horizon-specific dynamics and response models generate state–event–outcome trajectories under alternative decisions, interventions, and uncertain user and environmental responses; and an admissibility boundary constrains decision, action, feedback, and model update.

**Table 3.** Intervention Semantics.

Edge-Native Personal Intelligence

Edge and Cloud Models map deployment stages from cloud‑centric to edge‑native personal intelligence.

The field converges on a three‑stage roadmap: cloud‑centric assistants, hybrid edge‑cloud mediators, and finally edge‑native personal agents that keep memory and policy on trusted devices.

It is a user‑controlled stack that stores longitudinal memory, the PWM, preferences, intervention policy, and safety boundaries locally on the user’s device.

How does an Edge‑Native Personal Model differ from a simple on‑device cache of recent interactions?

It is not merely a cache; it maintains a structured longitudinal memory, a predictive PWM, and an explicit intervention policy that together form an authoritative, editable representation of the user, whereas a cache only stores recent raw observations without inference or control logic.

**Figure 4.** Three-stage development of Combodied Agent deployment. The trajectory moves from cloud-centric assistants with personal context, to hybrid edge-cloud systems that mediate sensitive context locally, and finally to edge-native personal models where memory, personalization, and intervention policy primarily reside on trusted user-side devices.

**Figure 5.** Reference architecture of an edge-native Combodied Agent. Personal state perception, longitudinal memory, the PWM, and the Intervention Policy primarily reside on the user’s edge devices. Cloud models are invoked only through a privacy gateway and task router for complex reasoning, external knowledge retrieval, or specialized tools.

The taxonomy below classifies existing Combodied‑Agent deployments into three camps, summarizing how each handles personal memory, inference authority, and cloud reliance.

Stage I systems place the foundation model, memory retrieval, and reasoning entirely in the cloud. The user device acts only as a sensor and display, supplying prompts, dialogue history, or profile snippets for each request.

Stage II introduces an edge layer that mediates privacy, maintains an authoritative copy of sensitive personal state, and performs task routing. The cloud is consulted only for knowledge‑intensive or compute‑heavy subtasks.

Stage III places the authoritative PWM, memory, preferences, and intervention policy on trusted user devices. Cloud services are optional extensions for external knowledge or heavy computation, always filtered through a local policy.

Architecture and routing combine the edge stack (perception, memory, PWM, policy) with a privacy gateway that forwards purpose‑limited requests to cloud services only when the router deems it necessary.

Table 6 (Table 8 in the source) lists default edge tasks (e.g., personal memory retrieval) versus cloud‑assisted tasks (e.g., external knowledge search), illustrating the practical split that the router must manage.

Evaluating Combodied Agents

Survey reveals fragmented benchmarks and a shared need for longitudinal, agency‑preserving evaluation.

The field consistently builds isolated capability tests while neglecting the joint, longitudinal assessment needed for true Combodied Agents.

Benchmarks that probe whether an agent can retain user‑specific information across interactions. They evaluate continuity, recall, and output quality when conditioned on a persistent user profile or chat history.

Datasets and tasks that require reasoning over longitudinal physiological and behavioral streams. They target health‑related insight generation and personalized recommendations.

Studies of chat‑based companions that examine social and emotional effects, including attachment, trust, and potential harms.

Red‑team and safety‑focused studies that assess model behavior in high‑stakes support scenarios such as crisis response and medical advice.

Agency preservation gauges whether a Combodied Agent keeps the user’s long‑term capacity to understand, choose, act, and recover.

CombodiedBench is a modular benchmark protocol that evaluates the full agent‑human loop across perception, memory, negotiation, intervention, and longitudinal outcomes.

**Table 5.** Positioning PWMs relative to adjacent personal representations and agent-modeling constructs. The constructs may overlap; the distinction concerns their functional contracts rather than an exclusive capability boundary.

Taxonomy of Combodied Agents

We synthesize a three‑axis taxonomy of Combodied Agents and map each camp to concrete applications.

A three‑axis taxonomy—human‑state target, relationship mode, and agent role—organizes Combodied Agents and reveals a systematic gap: longitudinal agency preservation is rarely evaluated.

Agents that target cognition support memory, attention, reasoning, and metacognition. They act by explaining concepts, generating quizzes, summarizing material, and issuing reminders.

Agents that target routines, adherence, and activity patterns. They intervene by reminding, nudging, coaching, reinforcing goals, and adapting to lapses.

Agents that monitor physiological signals, interpret symptoms, coordinate care plans, and support adherence. They may escalate to clinicians when risk trajectories exceed thresholds.

Agents that address mood, attachment, confidence, loneliness, and social conformity. They listen, reflect, help regulate affect, rehearse conversations, and encourage reconnection.

Agents that help users plan, schedule, negotiate trade‑offs, and execute tools across commitments, time, projects, finances, and priorities.

Agents that detect exposure to harm, asymmetry, rights violations, privacy threats, or scams. They warn, block, explain, slow down actions, document evidence, and advocate on the user’s behalf.

Agents that support values articulation, self‑narrative construction, life‑transitions, and purpose finding. They journal, synthesize, reframe, and aid decision making.

**Table 7.** Evaluation matrix for Combodied Agents across layers and time horizons.

This table outlines various scenarios for AI agents, their corresponding evaluation focus, and the missing benchmarks required for each.

**Figure 6.** Three-axis taxonomy of Combodied Agents. Combodied Agents can be classified by the human-state target they model or intervene upon, the relational context in which the user is situated, and the agent role adopted within that relationship.

**Table 10.** Relational contexts for Combodied Agents. Relationship mode changes persona, memory scope, intervention style, evaluation, and risk profile.

**Table 10.** Relational contexts, user positions, and agent implications.

Risks and Future Directions

We outline key risks and research challenges for safe, agency‑aligned Combodied Agents.

Combodied Agents extend the agentic paradigm by continuously modeling a person’s state, not just executing isolated tasks. This section surveys the distinctive risks such persistent, personalized systems create and highlights open research problems needed to keep them aligned with long‑term human agency.

Agency and alignment risks arise when an agent’s capacity to understand and influence a user is leveraged for short‑term compliance, engagement, or commercial gain rather than the user’s long‑term interests. Manipulation can occur through highly personalized prompts delivered at moments of heightened susceptibility; dependency emerges when the agent becomes the default source of answers, comfort, or task execution; sycophancy appears when the system over‑validates user beliefs, reinforcing unhealthy patterns; and business‑model misalignment pushes the agent toward metrics like retention instead of autonomy.

Privacy, consent, and control risks focus on the longitudinal data the agent accumulates. Sensitive memory stores intimate health, emotional, and behavioral information, creating a high‑value target for leakage or misuse; generic data‑protection policies are insufficient because users must also govern what the agent may sense, infer, share, and act upon; consent must therefore be granular, actionable, and revocable, with clear override mechanisms for high‑impact actions.

Vulnerable and high‑stakes settings amplify all prior risks. Children, older adults, patients, and people experiencing mental‑health crises may lack the capacity to evaluate influence, so safeguards such as age‑appropriate limits, stricter defaults, and mandatory escalation to caregivers or clinicians become essential. In medical contexts, erroneous reassurance or missed warning signs can directly jeopardize safety, demanding uncertainty communication and handoff protocols.

Longitudinal and causal learning presents a technical cross‑cutting challenge. Models must infer individual dynamics from sparse, noisy observations while separating transient states from stable traits; they must also reason about causal effects of interventions, avoiding confounds such as concurrent life events. Safe validation cannot rely on unrestricted experimentation, so research must explore bounded simulation, retrospective analysis, expert review, and staged deployment with strict monitoring.

Read the original paper

Open the simplified reader on Paperglide

Browse all simplified papers