Feng Bingqing, CTO of Huajiao Live and founder and CEO of Chaoxu AI, led his team to begin developing an enterprise‑grade Agent platform in March 2026. Within one month, they launched the first minimum viable version, featuring Agent management, Skill configuration, access control, and chat‑interaction capabilities. As of October 26, 2026, approximately 30 Agent applications have been deployed across Huajiao’s business lines and within the Huafang Group’s mid‑end platform, covering a range of real‑world use cases such as R&D automation, operational data querying, customer service response, preliminary legal review of contracts, administrative task processing, and government relations (GR) support.
The platform adopts a clear four-tier division of responsibilities: the business entry point handles task submission and result presentation; the control and governance layer centrally manages identities, permissions, configurations, versions, and usage metrics; the Agent Runtime is responsible for model inference, tool invocation, and state progression; and the resource integration and execution environment connects to the enterprise’s existing systems to carry out specific tasks. The platform adheres to the principle that “the common kernel does not encapsulate department-specific logic,” with business-specific differences expressed through Skills, Prompts, tool interfaces, and configuration settings, ensuring that new scenarios can be added without modifying the shared runtime kernel.
To address the challenge of isolation in a multi-user shared execution environment, after evaluating Docker, virtual machines, and MicroVM solutions, Huajiao selected Tencent Cloud’s open-source project CubeSandbox—due to its fast startup, strong isolation, and performance that meets requirements. This solution has been running stably in its production environment for over six months, supporting approximately 30 agents across two instances, Huajiao and Huafang, with an average daily token consumption of about 83 million. Huajiao has also developed its own terminal‑protocol agent, agentd, to interface with CubeSandbox, resolving key issues such as environment variable propagation, bearer‑token authentication, and file visibility verification.
At the organizational collaboration level, Huajiao adopts a “unified public code, separate organizational configuration” model to enable cross-departmental reuse: Huajiao and Huafang share the same Agent Runtime codebase but deploy independent instances, with their account systems, business data, and execution environments fully isolated. At this stage, the focus remains on improving efficiency at individual touchpoints; in the next phase, two key directions will be prioritized: first, integrating distributed Agents into end-to-end business workflows (such as product requirements → design → development → testing for R&D automation); second, reverse-engineering Agent requirements from the group’s operational objectives (e.g., user engagement, room‑entry rate, live‑stream conversion) to facilitate cross‑departmental data sharing and joint analysis within authorized scopes.
Huajiao emphasizes that AI value assessment should return to real-world task costs: it must account not only for model‑call expenses, but also for human review, rework, ongoing maintenance, and the risk of failure. There was once a reconciliation‑type agent that consumed tens of thousands of yuan per user per day in production, prompting the team to establish technical selection guidelines: “prioritize deterministic tasks; delegate judgment‑heavy portions to models; and retain human oversight for critical responsibility points.” Feng Bingqing points out that the core competitive barrier for enterprise AI lies not in the model or framework, but in the organization’s accumulated business rules, domain expertise, system interfaces, and data permissions—elements that cannot be purchased externally and must be internalized through continuous practice to become organizational capabilities.