火山引擎撕开“Agent元年”真相：大模型已演进到按使用量订阅

Core Insights - The large model industry is entering the "Agent Year," with significant transformations occurring in the enterprise market, led by Volcano Engine's HiAgent intelligent platform, which has achieved the highest rankings in both contract amounts and numbers [2] - Industry data indicates that by the first half of 2025, the number of projects awarded for intelligent agent construction platforms reached 371, representing a 250% year-on-year increase, nearing two-thirds of the total from the previous year [2] - Volcano Engine has maintained its position as the leader in the number of contracts since the second half of 2024, with a strong growth momentum reflected in its dual championship in contract amounts and numbers in the first half of this year [2] Company Strategy - Volcano Engine's Vice President, Zhang Xin, emphasized that enterprises require definitive solutions that transform industry know-how into productivity rather than random technological attempts [2] - The HiAgent 2.0 platform features four core upgrades: the innovative Agent DevOps system for integrated development and operations, an industry template library that reduces customization cycles by 80%, a unified model application system for fine-tuning exclusive models, and the Canvas interactive portal for centralized management of hundreds of dispersed agents [2] Industry Trends - The business model in the industry is undergoing a fundamental shift, moving from pay-per-use and token-based pricing to a subscription model based on usage, with the ultimate goal being a revenue-sharing model focused on "marketing revenue and cost savings" [3] - Addressing concerns that agents do not enhance cloud vendor stickiness, Zhang Xin argued that the long-term memory and knowledge base of agents create natural barriers, with Volcano Engine building a cost-effective moat through its Doubao large model and Volcano Ark inference optimization [3] - A case was cited where replacing HiAgent allowed for a breakthrough in concurrent capabilities, exceeding the previous platform's limit of 200, thus meeting real production demands [3]