What's changed: from reactive tools to proactive partners
The old view of AI assistants was simple: you ask, they answer or act—like sending a reminder or fetching a file. That's being overturned by the idea of combodied agents, which treat your evolving state (health, mood, intentions) as the primary focus, not just the task at hand [1]. Instead of a one-off reminder, a combodied agent would notice you missed a dose, consider why (forgot, confused, side effects), and then choose an action—like bringing the medication or contacting a caregiver—based on your consent and safety [1]. This is a fundamental shift from 'task completion' to 'sustained human benefit.'
This shift is supported by a concrete prototype: the Personal Health Agent (PHA) built by researchers at Google and universities, which uses a multi-agent framework to handle three distinct roles—data analysis, domain expertise, and coaching—so it can reason about your wearable data and health records to give personalized advice [3]. They tested it across 10 benchmark tasks with over 7,000 annotations and 1,100 hours of expert and user effort, making it the most comprehensive evaluation of a health agent to date [3]. That scale shows the field is moving from theory to serious validation.
What combodied agents will actually do for you in the next two years
In the near term, expect personal assistants to become more context-aware and proactive. Instead of just setting a timer, they'll track your patterns (e.g., sleep, activity) and suggest adjustments before you ask [1][3]. For health, the big win is personalized, adaptive interventions: a platform described in a 2024 paper uses AI and reinforcement learning to send tailored recommendations based on your past data and real-time monitoring, which could significantly improve health outcomes, especially in resource-poor settings [2]. That means your health agent could adjust its coaching based on your progress, not just follow a static plan.
For companions, the change is about emotional and social intelligence. Combodied agents are designed to model your 'personal world'—your likely future states under different decisions—so they can offer support that feels proportionate and respectful of your autonomy [1]. For example, if you're feeling down, a companion agent might suggest a walk or a call to a friend, but only if it judges that's what you'd want, based on your history and preferences [1]. This is a far cry from today's chatbots that respond to prompts without understanding your deeper context.
The catch: what still needs to improve before you can trust them
The biggest hurdle is trust and safety. Combodied agents will have access to sensitive data and the ability to act in the physical world, so they must be designed with consent, uncertainty, and reversibility in mind [1]. The framework explicitly calls for 'admissible intervention policies' that respect user control—meaning the agent should ask before acting, and you should always be able to override it [1]. That's a tall order, and the papers are clear that these are early-stage concepts, not yet deployed at scale.
Another challenge is integration. The Personal Health Agent [3] and the adaptive intervention platform [2] both rely on pulling data from multiple sources—wearables, health records, and user input—and making sense of it all. That's technically hard, and the papers note that current systems are fragmented across tasks and modalities [4]. So while the vision is compelling, the next two years will likely see incremental progress: better integration, more robust evaluation, and pilot deployments, rather than a seamless all-in-one agent. The evidence is strong that the direction is right, but the road is still being built.
About These Sources
This answer is built on 4 studies (all preprints) — published from 2024 to 2026, 4 from 2024 or later — selected as the most relevant from 4 studies that passed quality screening, drawn from 36 papers retrieved from a database of over 500 million.
Sources used in this answer
ComBodied Agents: a New Paradigm of Human-Centric Agentic AI
Introduces the combodied agent paradigm, which models and supports individual human-state trajectories over time, using a closed loop of perception, memory, prediction, and intervention, with explicit emphasis on consent and user control.
The Digital Transformation in Health: How AI Can Improve the Performance of Health Systems
Describes an AI and reinforcement learning platform that delivers adaptive health interventions, integrating multiple data sources and digital health apps, with potential to significantly improve outcomes, especially in resource-poor settings.
The Anatomy of a Personal Health Agent
Presents the Personal Health Agent (PHA), a multi-agent framework for personalized health recommendations, evaluated across 10 benchmark tasks with over 7,000 annotations and 1,100 hours of expert and user effort, making it the most comprehensive health agent evaluation to date.
Human-Centric Intelligence in the Era of Foundation Models: A Survey
Surveys human-centric intelligence in the foundation-model era, proposing a taxonomy of six levels from visual appearance to embodied agency, and highlights the fragmentation across tasks and modalities as a key challenge.