AsianFin— The rise of AI Agents is the headline story of 2025. Yet, despite the buzz, the industry faces a sobering reality: large models have landed, but they have yet to fully take root.
On the surface, 2025 has been branded the “Year of the Agent.” Viral consumer platforms like Manus and a flood of AI Agent showcases at this year’s World Artificial Intelligence Conference (WAIC) have propelled Agents to the forefront of public and industry attention.
While consumer-facing applications continue to capture imaginations, enterprise-grade Agents—designed to address real business needs and drive revenue—are inching closer to commercial reality. The flourishing popularity of Agents reflects the broader maturation of large model applications.
However, behind the scenes, many enterprises remain bogged down in the complexities of implementation. High infrastructure costs, entrenched data silos, and elusive business value have left many companies navigating through a frosted glass, where flashy exhibition demos obscure the messy realities of day-to-day operations. Much-anticipated disruptive applications are languishing in prolonged proof-of-concept stages, awaiting a more pragmatic route to scalability.
“No enterprise wants to flip a coin and gamble on uncertain outcomes,” said Zhang Xin, Vice President of Volcano Engine, in a recent interview with AsianFin. “What businesses need is a clear path to converting their industry expertise into tangible productivity gains through large models.”
Despite these hurdles, the market is undeniably heating up. In 2024, there were 570 contract-winning projects linked to intelligent agent platforms, with disclosed contract values totaling 2.352 billion yuan. The first half of 2025 has already seen 371 project wins—3.5 times the volume of the same period last year, and nearly two-thirds of 2024’s total, with demand expected to surge further in the second half.
Volcano Engine, a subsidiary of ByteDance, has emerged as a dominant player. Since the second half of 2024, it has consistently topped the charts in both contract value and volume, thanks to its full-stack intelligent agent platform, HiAgent. According to Chen Xi, Head of HiAgent, success in enterprise AI demands far more than just a strong model; it requires a tightly integrated solution blending technical tooling, business adaptation, security, services, and proven best practices.
“Having a great model doesn’t automatically translate to great applications,” Zhang said. “The missing link is robust engineering practices—prompt design, orchestration, privacy controls, system integration—all wrapped into a development platform like HiAgent.”
Volcano Engine’s approach reflects a broader industry shift. In the early days, companies focused heavily on improving LLM capabilities, believing better models would naturally lead to better applications. However, as Zhang noted, the realization has set in that models are just one piece of a complex puzzle. The real challenge lies in building scalable platforms that can convert model potential into business-ready solutions.
Platforms like HiAgent have evolved into end-to-end workspaces for enterprise AI Agents. Beyond development, these platforms now manage the entire lifecycle—planning, deployment, monitoring, and optimization—mirroring the DevOps methodologies that transformed cloud-native application development.
One of HiAgent’s newest features, the Data Flow Module, continuously feeds back data from agents’ operations, enabling dynamic learning and iterative improvements. “An Agent isn’t a static product,” Chen emphasized. “It needs to evolve, learning from real-world usage just like a human employee.”
Even as technological foundations mature, enterprises still struggle with operational challenges. Fragmented use cases, inconsistent deployment standards, and a lack of agent governance frameworks hinder scalability. HiAgent’s latest update introduces the Canvas Interactive Portal, a unified entry point that streamlines access to hundreds of in-house Agents, allowing employees to quickly find and deploy digital counterparts across business functions.
Despite the rapid growth of Agents, Zhang remains cautious. “We’re in the divergent stage of intelligent agent platforms, similar to the early 'battle of a hundred models' before the MoE architecture became mainstream,” he said. The ecosystem still lacks widely accepted agent maturity frameworks and standardized best practices.
From a commercial perspective, Zhang believes the business model for Agents must pivot toward outcome-based pricing. “The early phases were all about selling compute time or tokens. But for Agents to scale in enterprise environments, pricing needs to reflect real business value—like how much revenue they generate or operational costs they save.”
IDC predicts that through 2025, generative AI adoption will remain anchored in productivity applications—office assistants, CRM upgrades, and industry-specific workflows in finance, energy, and manufacturing. Agents represent the next frontier, poised to revolutionize process automation and digital workforce strategies.
However, Zhang cautions against oversimplifying the path forward. “There are general-purpose large models, but it’s incredibly difficult to create general-purpose Agents,” he said. Agents are inherently scenario-dependent, requiring customization and continuous tuning. Moreover, enterprise leadership often overestimates their immediate potential, while frontline users underestimate their practical utility.
HiAgent aims to address this by lowering barriers across the enterprise AI value chain, from simplifying agent development with industry templates to offering integrated model-level toolchains for clients seeking domain-specific customizations.
Volcano Engine also sees a critical opportunity in bridging the “last mile” of AI adoption. The company is expanding its industry-specific sample rooms and template libraries, targeting common use cases in customer service, marketing, HR, and office administration. These are further segmented into verticals like healthcare, education, and finance, allowing enterprises to leap from “0.8 to 1” with minimal customization.
Zhang emphasized the long-term vision: “No company is paying for ‘the TCP/IP protocol,’ but they will pay for applications built on it. For Agents, true prosperity will come at the application layer—when they’re seamlessly integrated into business processes, absorbing the capabilities of large models underneath.”
At the infrastructure level, Volcano Engine continues to invest in optimizing inference performance and cost-effectiveness. ByteDance’s Doubao large model initiative and Volcano Ark’s AI acceleration strategies reflect this dual focus. The goal is to ensure that Agents aren’t just technically feasible, but economically viable for enterprise-scale deployment.
In Zhang’s view, unlike the era of cloud-native containers that decoupled developers from cloud vendors, Agents create inherent stickiness. “Agents accumulate knowledge and long-term memory. The more a company uses them, the better they get. If that data isn’t portable, migrating platforms becomes impractical,” he said.
The “Aha moment” for human-machine collaboration, Zhang recalled, came when Guangzhou Public Transport Group issued an employee ID to a digital Agent, fully integrating it into their workforce for contract reviews and vehicle maintenance. This, he said, is the real sign that Agents are moving from demos to production lines.
Yet, challenges remain. For Agents to move from 2025’s hype to long-term enterprise transformation, platforms, ecosystems, and business models must evolve in tandem. The industry’s future hinges not just on better models, but on delivering end-to-end value that enterprises can measure, scale, and trust.
更多精彩内容，关注钛媒体微信号（ID：taimeiti），或者下载钛媒体App
