Building Human Centric Agentic AI Systems

Summary & Methodology Analysis

The proposed architecture departs from existing digital or embodied agent paradigms by centering the system on the human subject. The methodology initiates with event based multimodal perception to reconstruct personal history, which is then stored in a longitudinal, correctable memory structure. This data feeds into Personal World Models, which are probabilistic frameworks that output calibrated distributions over potential future states and outcomes, allowing the agent to reason about the long term trajectory of the user. The system relies on a continuous feedback loop where human and environmental responses update the perception and memory cycles, ensuring the agent adapts to evolving needs.

To manage intervention, the system utilizes an admissible policy that selects actions subject to strict constraints. These constraints include explicit user consent, safety protocols, reversibility, and mechanisms for human control, effectively mitigating the risks associated with automated decision making. The framework integrates numerous data sources and models such as OpenFace, Ego4D, Health-LLM, PHIA, OPPORTUNITY, DeepConvLSTM, AWARE, StudentLife, Sotopia, MIMIC-IV, Med-PaLM, BEHRT, Med-BERT, MemoryBank, LongMemEval, GeMAPS, eGeMAPS, and MERBench to handle the diverse multimodal inputs required for high fidelity personal tracking.

Despite the architectural advancements, the paper acknowledges a primary limitation in the current state of technology. The creation of a holistic, high fidelity digital twin that represents the totality of a human person remains infeasible with current methods. Consequently, while the framework provides a robust structure for agentic support, the depth of personal modeling is bounded by current representational limits and the inherent complexity of mapping human life trajectories.