From Language Models to World-Acting Systems: Progress and Limits of Agentic AI across Digital, Social, Virtual, and Physical Environments

As of August 31, 2026, large language models have evolved into consequential agents that allow for changes to external states. Existing research evidence shows that action interfaces are the most fully developed, but there is insufficient evidence in terms of robust completion, recovery, authorization, and independent verification. Although the Model Context Protocol and Agent2Agent have improved interoperability, a reliable delegation mechanism has not been established; multi-agent organizations have increased specialization levels, but this has also led to increased costs and related failure risks. Persistent simulation and world models support training planning, but they have not proven to be truly agency-efficient. Robotic and autonomous driving laboratories have only established limited feasibility, not the reliability of operations in open-world scenarios without humans. It is recommended to use justified delegation as an analytical criterion, with evidence supporting traceability, boundary authorization, failure detection, safe recovery, and calibration…