71% of surveyed enterprises identify performance evaluation and go-live acceptance as the leading challenges in deploying agents, while 43% focus on multi-agent collaboration and system integration, and 41% care about cost attribution and return on investment. These findings show that enterprises’ enthusiasm for agents—AI programs that autonomously call tools and complete multistep tasks—is giving way to concerns about operational quality and returns.
Alibaba Cloud launched the fifth stop of its AgentCore eight-city salon tour in Nanjing on November 20, aiming to help enterprises move from experimental projects to production through a platform spanning agent creation, hosting, observability, security auditing, cost management and model routing. The emphasis is not on adding another question-and-answer interface, but on connecting agents to enterprise operations and keeping them running.
What this is
AgentCore is Alibaba Cloud’s enterprise platform for building and governing agents. The Nanjing event covered five main areas: creating or bringing existing agents under management; administering identity, access permissions and security audits; continuously evaluating performance; using STAROps for data analysis and troubleshooting; and selecting the right model for different tasks through an AI Gateway.
At its core, this is an enterprise-grade operating framework rather than a single model. Multi-agent collaboration means multiple agents with different responsibilities working together, while model routing dynamically selects models and call paths according to task requirements. The event also offered hands-on exercises, with participants required to bring their own laptops to build or manage an agent.
Industry view
Alibaba Cloud has identified one of the most practical needs in enterprise deployment: pilots may be owned by R&D teams, but moving them into production brings permissions, data, operations, cost and accountability into play. The three leading problems identified in the Hangzhou survey closely match the platform’s selling points, showing that the vendor clearly recognizes the barriers facing adoption.
However, the approach is not yet sufficient to demonstrate value at scale. First, the data came from participants at a vendor-organized event; the sample scope and statistical methodology were not disclosed, so the 71% figure should not be generalized to all enterprises. Second, multi-agent collaboration, observability and governance also increase system complexity: platform capabilities do not automatically translate into business returns. Third, customer cases and hands-on certificates presented at the event mainly validate product workflows; they cannot replace long-term testing of failure rates, human-intervention ratios and cost per task.
Impact on regular people
For enterprise IT: When enterprises procure agent services, their evaluation will shift from “How capable is the model?” toward “How are permissions managed, how is performance validated, and how are costs calculated?” Existing systems, identity frameworks and operations processes will become the practical barriers determining whether a project reaches production.
For individual professionals: Agents will increasingly enter process-driven work such as data analysis, inspections and log analysis, but performance acceptance will become more quantitative. Whether a role’s tasks can be completed reliably—not merely answered impressively every so often—will influence how those tasks are reassigned.
For the consumer market: Intelligent customer service, order processing and service scheduling in retail, transportation and consumer services may gradually move from individual models to uniformly governed multi-agent systems. Consumers may never see the platform, but they could experience more coherent services or stricter permission and audit mechanisms.