(Yicai) July 17 -- The global technology industry is at a critical transition from generative artificial intelligence tools to the agent era this year. Software is no longer a passive auxiliary tool that merely responds to commands but has evolved into an economic entity with proactive intent and the ability to recursively execute complex tasks.

This transformation is first reflected in the reshaping of infrastructure. Traditional enterprise software architectures are based on the 1:1 human-machine interaction assumption, according to Silicon Valley-based venture capital firm Andreessen Horowitz, better known as a16z. But the recursive operations of AI agents can instantly break down a single instruction into tens of thousands of concurrent tasks, generating a massive number of application programming interface calls.

This sudden high concurrent traffic subjects traditional systems to the "thundering herd effect," forcing infrastructure to drastically reduce latency variance and shift the bottleneck from computing power to the coordination layer's routing and distribution capabilities, a16z noted. 

The data layer also faces challenges. Eighty percent of a company's high-value knowledge is not stored in structured databases but is scattered in unstructured formats, such as contracts of a PDF, meeting videos, Slack chat records, and system logs, said Jennifer Li, a partner at a16z. The sensitivity to the quality of data input increases exponentially as models become more intelligent.

The opportunity this year lies in the semantic structure of multimodal data, truth arbitration, and timeliness governance to prevent AI agents from making wrong decisions based on outdated or conflicting information. Video technology has evolved from a media format to a world model that understands physical laws, serving as simulation environments for training of robots and autonomous driving systems. 

The power structure of the enterprise software market is being inverted. AI agents can directly access and manipulate underlying data to perform tasks, bypassing the cumbersome graphic interfaces. Traditional record-keeping systems, such as Salesforce and SAP, are degrading into mere data storage, with value shifting to 'action systems' that can complete actual business tasks for users.

The business model is also transitioning to an economy that charges based on results and task complexity from one based on seat fees and screen time. AI applications in vertical fields are beginning to build unified agent ecosystems, enabling different AI agents representing different parties to directly collaborate and negotiate with each other. 

The physical world is undergoing a synchronized software-driven reinvention. The manufacturing thinking is becoming 'API-ized,' with future factories becoming replicable standard modules themselves.

Software designs production enzymes through directed evolution algorithms, optimizes nuclear reactor structures, and shifts from 'devouring the world' to 'moving the world.' Billions of connected sensors make observing the entire physical world in real time possible, and enterprises build competitive barriers by capturing proprietary physical operation data.

This has given rise to new business models: the health industry shifts from treating sick people to maintaining healthy people, and education moves from memory-based assessment to evaluating the ability to solve problems using AI. 

A clear thread runs through this year: autonomy. Infrastructure expands the capacity of machine autonomous recursion, the software economy reorganizes due to agent autonomy, and the physical world is software-defined through autonomous factories. The core value of the future will belong to those who build control planes, trust protocols, and physical interfaces for AI agents.

The authors are Yang Yanqing, director of the Center for Education, Innovation, and Sustainable Development Studies at ShanghaiTech University, and An Xu, an AI observer.

Editor: Futura Costaglione