Large model competitions are moving from "capability contests" to "practical application challenges." At the recent Volcano Engine Force Original Power Conference, Volcano Engine President Tan Dai systematically outlined a new paradigm in AI evolution: Intelligent Agents (Intelligent Entities) will become the core carrier for AI implementation. Multi-modal capabilities and an efficient Agent development system are the key to bridging the gap between technology and industry.
From "chatting" to "working": Large models enter a phase of tackling complex scenarios
Tan Dai pointed out that in the past, large models were mainly used for question-and-answer interactions. Now, they have penetrated high-complexity industries such as automotive, manufacturing, and catering. In these scenarios, AI needs to process text instructions, visual input, sensor data, and tool outputs simultaneously. For example, identifying equipment abnormalities in a factory and calling a maintenance ticket system, or generating nutritional analysis and recommendations based on images of dishes in a restaurant. This requires models to have human-like multi-modal understanding and environmental operation capabilities, rather than relying solely on pre-defined APIs.
Agent development becomes the biggest bottleneck, Volcano Engine launches AgentKit to break through
"The model's capabilities are already strong enough, but how to package them into stable and scalable agents remains a major industry bottleneck," Tan Dai admitted. To address this, Volcano Engine officially launched AgentKit—a smart agent development and operation framework derived from internal practices, providing full-chain components such as task planning, tool calling, memory management, secure sandboxing, and monitoring and tracking, significantly lowering the development and maintenance costs of agents.
Agents will become the "new computing unit" in the AI era
Tan Dai further predicted that the core infrastructure of the AI era will shift from web pages and mobile apps to intelligent agents. This means cloud architecture must be restructured—databases need to support agent state persistence, computing resources need to be dynamically scheduled according to tasks, and networks must ensure low-latency communication for multi-agent collaboration. "An agent is not a functional module, but a digital employee with goals, memory, and the ability to act," he said.
Safety must be inherently embedded in Agent design
Facing the risks of AI abuse, Tan Dai emphasized that traditional boundary protection has failed, and safety capabilities must be deeply integrated throughout the entire lifecycle of Agent operations. Volcano Engine has already integrated mechanisms such as input filtering, output compliance verification, approval for sensitive operations, and behavior auditing into AgentKit, ensuring reliable operation of Agents in open environments.
AIbase believes that Volcano Engine's latest release marks the transition of domestic large model vendors from "model suppliers" to "builders of intelligent agent operating systems." When AI no longer just answers questions but actively performs tasks, true industrial intelligence truly begins. The open-source nature and cloud-native integration of AgentKit may become a key accelerator for Chinese companies embracing the "Agent economy."
