Abstract: AI agents are rapidly transforming how tasks are delegated and executed in digital environments. Unlike traditional generative AI systems that mainly produce content, AI agents autonomously decompose goals, call tools and act across applications, turning into task proxies and transaction gateways. This evolution fundamentally challenges the traditional logic of authorisation, i.e., user's preference, willingness and authorisation are expressed at the front end, while decisions and actions unfold through a long, opaque chain, occasionally producing legally relevant effects in the external world. In China, the recent guidance on AI agent issued in May 2026 explicitly foregrounds user authorisation scope, decision-making boundaries, traceability and users' ultimate decision-making rights, yet current rules remain largely principled and do not resolve how to allocate liability within complex authorisation relationships. This paper provides a typology of authorisation and reconstructs the background assumptions that underpin traditional authorisation mechanisms. On this basis, it further demonstrates how, in AI-agent settings, authorisation media become insufficient, linguistic expressions are compressed and expanded by natural-language interaction, subjects of authorisation become obscure and the consequences of authorisation become increasingly unpredictable. As a response, the paper proposes a governance framework for AI-agent authorisation, and the central claim is that authorisation should be reconceived from a one-off proof into a dynamic governance mechanism spanning ex ante, in-process and ex post stages, thereby enabling innovation under conditions of legally manageable risk.
Keywords: AI agents; authorisation; agency; liability allocation; agent economy
Author: Li Wenlong
Source: Li, W. L. (2026). Reconstructing Authorisation in the Age of AI Agents. Legal Forum, 41(4), 66–79.
Publish date: 2026/07/03
Original Link: 
智能体时代授权的逻辑限度与法律建构_李汶龙.pdf
