AI Agents Need More Than Models: The Digital Infrastructure Gap

Introduction: Beyond Model-Centric AI

Discussions about artificial intelligence over the past year have fixated on large models—bigger models, more agents, finer prompts. While not wrong, this view only sees the surface. As AI shifts from chat tools to task-executing systems, what determines real-world deployment is not a single model's parameters but the underlying digital foundation: stable network connectivity, nearby compute scheduling, compliant data circulation, controllable security, and reorganized business processes. In short, the agent era competes on the coordination of cloud, network, edge, end, data, and security.

Policy Signal: From Point Applications to Infrastructure Coordination

China's Ministry of Industry and Information Technology (MIIT) recently issued the "Artificial Intelligence + Information Communication Innovation Development Implementation Opinions (2026–2028)." Key targets for 2028 include:

Preliminary formation of a mutually reinforcing innovation landscape between AI and information communication.

30+ high-value typical scenarios and a batch of characteristic intelligent agents.

Enhanced infrastructure support: metropolitan-area compute 1-millisecond latency circle coverage ≥ 75%.

These goals signal a clear transition: AI is moving from "single-point applications" to "infrastructure coordination." Earlier, an AI app could sit in the cloud, accept a question, and return an answer. Today, scenarios like urban incident detection, factory equipment inspection, traffic anomaly identification, government service assistance, and emergency multi-source information fusion require communication networks, compute platforms, edge nodes, business systems, data resources, and security mechanisms to work together. The deeper AI penetrates the physical world, the higher the demand on the foundation.

Agents Are Not Better Chatbots—They Must Plug Into Processes

Many equate agents with auto-task-decomposing chatbots. Valuable agents, however, must enter business workflows. The article defines four essential capabilities:

Business context understanding – Not just a single sentence, but knowing which process a task belongs to, the rules involved, required data, and who consumes the output.

Tool invocation – Querying knowledge bases, generating documents, calling APIs, reading spreadsheets, retrieving logs, triggering approvals, producing reports.

System collaboration – Integrating with existing business systems rather than rebuilding everything on a new platform.

Auditability and constraint – Traceable task initiation, data access, result generation, and human confirmation.