Combodied Agents are a proposed third paradigm of agentic AI whose primary target is not software state or physical state but the evolving human, closing a loop over multimodal perception, longitudinal memory, a Personal World Model (PWM), and consent-bounded intervention while measuring success by preserved human agency, not task completion.
You’re building an assistant that texts an older user medication reminders, pulls heart-rate off their watch, and can dispatch a caregiver. Today your success metric is “did the pill get taken.” But that metric can’t tell you if the user forgot, refused, is confused, or is drifting into over-reliance on the bot. This paper argues that both dominant agent styles miss this. Digital Agents optimize software state (WebArena, SWE-bench); Embodied Agents optimize physical state (PaLM-E, robotic control). Neither treats the person’s trajectory over weeks and months as the thing being modeled and evaluated.
The paper is a position and framework paper, not a system with benchmarks. It defines a Combodied Agent by its action substrate: the class of states that organizes modeling and evaluation. Digital agents act on digital states, embodied agents on physical states, Combodied Agents on evolving human states plus agency. The contribution is a closed loop with four functional blocks. First, event-based multimodal perception turns sparse, noisy signals (text, speech, vision, wearables, motion, social, environmental, clinical records) into governed event-evidence records that carry provenance, uncertainty, and consent scope. Second, longitudinal memory stores events, goals, interventions, outcomes, and corrections, and must be user-inspectable and deletable. Third, a Personal World Model (PWM) predicts distributions over future human states and outcomes conditioned on candidate interventions, not just next-user-behavior. Fourth, an admissible intervention policy selects actions only within a set constrained by consent, reversibility, uncertainty, safety, and escalation. Non-intervention, clarification, and referral are first-class actions.
The authors distinguish this from a Human Digital Twin: rather than a high-fidelity replica of the whole person, a Combodied Agent keeps purpose-bounded, correctable slices. They also stage deployment from cloud-centric, to edge-mediated, to edge-native, where authoritative memory and intervention authority sit on user-controlled devices.
while True:
obs = sense(modalities) # text, audio, vision, wearables, records
events = to_event_evidence(obs, provenance, uncertainty)
memory.update(events, consent_scope)
z = infer_state(memory, context, goals) # uncertainty-bearing posterior
trajectories = pwm.rollout(z, candidate_interventions, context)
admissible = filter(candidate_interventions, consent, safety, reversibility)
action = policy.select(admissible, trajectories, agency_objectives)
outcome, feedback = act_or_abstain(action) # includes clarify / escalate / do nothing
memory.log(action, outcome, feedback)
The prevailing default in agent design is that success equals task completion with less human involvement: longer horizons, more autonomy, higher scores on benchmarks like SWE-bench or OSWorld. This paper inverts that. The primary object of modeling and evaluation should be the person’s trajectory and preserved agency, and the agent’s tools, sensors, and robots are just channels toward that. The forward-bridge to evidence is that the authors reorganize benchmarks, taxonomy, and deployment around this reframing rather than proposing a new model.
This is a synthesis paper, so there are no experimental numbers. The load-bearing claims are conceptual:
•
Existing agent categories cover fragments, not the loop. MemGPT-style memory adds recall but not intervention response; Health-LLM reasons over wearables but doesn’t validate longitudinal outcomes; companion systems affect loneliness and attachment but lack relationship-safety benchmarks. No category integrates perception, memory, prediction, intervention, and agency evaluation.
•
A Personal World Model (PWM) is defined by an intervention-conditioned contract, not by personalization or plausible generation. Predicting the next user message doesn’t count. Predicting how this person’s state and outcomes shift under alternative agent actions does.
•
Agency preservation is proposed as a first-class, multi-dimensional metric family, with sub-metrics for autonomy, contestability, informed choice, capability growth, over-reliance, reversibility, boundary respect, and preserved human relationships. High engagement should not offset failures on these.
•
The authors propose CombodiedBench as a suite spanning human-state perception, memory continuity, goal negotiation, intervention appropriateness, agency preservation, relationship boundaries, escalation, and longitudinal outcomes. It is described as a proposal, not a released benchmark.
•
A three-stage deployment path (cloud, hybrid, edge-native) argues that authoritative personal memory and intervention authority should migrate toward user-controlled devices as edge capability allows.
Reach for this framing when you’re designing a health, learning, eldercare, or companion product where the same user comes back for months. Instead of logging “reminder sent, task completed,” instrument the loop the paper describes: an event log with provenance and consent, an explicit personal state estimate with uncertainty, a prediction step that scores candidate interventions (including do nothing and ask a human), and metrics for capability growth and over-reliance alongside task success. The intervention taxonomy (inform, remind, coach, coordinate, execute, escalate) with reversibility flags is a practical checklist for a product spec.
The paper releases no code, models, or datasets. CombodiedBench is proposed rather than built. Related resources cited that you can actually use include LongMemEval for long-horizon memory, Ego4D for first-person perception, and public wearable-reasoning work like Health-LLM. Treat the paper as a design vocabulary and evaluation checklist, not a stack.
Make the person, not the task, the thing your agent is optimizing, and measure whether repeated use leaves the user more capable or more dependent. Task completion and engagement are cheap to measure and easy to game. Agency, capability growth, and appropriate escalation are the metrics that separate a useful long-term companion from a well-tuned dopamine loop.
•
No empirical validation. The framework, the Personal World Model (PWM) contract, and CombodiedBench are proposals. Nothing here has been shown to work end-to-end on real users.
•
Agency-preservation metrics are defined qualitatively. Operationalizing “capability preservation” or “over-reliance” into scorable signals in a shipping product is unsolved and will vary by domain.
•
The edge-native stage assumes local models capable enough to host memory, a personal world model, and policy on user devices. For most current stacks that means degraded reasoning versus cloud, and the paper doesn’t quantify that trade-off.