From Single‑point Copilot to Platform‑level Agentic: Real Challenges and Future Paths for Data Platforms

The DataFunSummit livestream titled “Agentic Data Platform Challenges and Future” brought together host Yi Long with data experts Wang Tonghuan (vivo) and Wang Bo (YangQianGuan) to examine how data platforms can evolve from a single‑point Copilot to a platform‑level Agentic system.

Defining Platform‑level Agentic – The core question was not whether AI can write SQL, but whether it can continuously, reliably, and responsibly generate and execute tasks in complex enterprise environments. Wang Tonghuan emphasized that Copilot keeps humans at the center while Agentic systems aim to automate goal‑oriented planning, tool invocation, and result verification, gradually shifting the AI from a passive assistant to a central orchestrator.

New Infrastructure Requirements – Both speakers agreed that a true Agentic platform must add several previously hidden modules: short‑term and long‑term memory, planning capabilities, a stable tool bus, dynamic context budgeting, and comprehensive logging. Wang Bo described the platform as an “operation plane” composed of an orchestrator, knowledge system, execution environment, and observation system, rather than a single model call.

Security Guardrails – In highly regulated financial scenarios, Wang Bo highlighted the need for pre‑emptive intent detection, whitelist filtering, and framework‑level hooks to block dangerous DDL/DML actions. Wang Tonghuan added process‑level controls such as read‑only accounts, gray‑scale table writes, and human review at critical nodes, stressing that the risk is not only malicious misuse but also inaccurate outputs that could mislead business decisions.

Knowledge Base as an Enterprise Semantic Layer – The discussion shifted to the importance of a knowledge repository that serves Agents directly. It must include metric definitions, table lineage, typical queries, error cases, permission rules, and tacit expertise, all structured for machine consumption. This transforms the semantic layer from a documentation aid into a core interaction interface.

Evaluation Failures – The first round of testing often falls short not because the model is weak, but because multiple engineering layers—retrieval, context assembly, tool execution, and task planning—break down together. Participants stressed the need to pinpoint which module fails, capture the failure for knowledge‑base updates, and treat evaluation as an engineering problem rather than a pure model‑accuracy issue.

Stability and Observability – Introducing Agents adds uncertainty, requiring new stability mechanisms: resource isolation, gray‑scale execution, read‑only runs, and human checkpoints. Wang Bo emphasized full traceability of each decision step to enable root‑cause analysis, while Wang Tonghuan noted that monitoring must now cover the Agent’s reasoning path, not just system metrics.

Measuring Success – Adoption metrics (clicks, likes) are insufficient. The panel agreed that true success is measured by task‑completion quality, net promoter score, and concrete business impact such as improved insight generation or sales uplift. Evaluation should cover syntax correctness, result consistency, intent alignment, and AI quality scores.

Cost and ROI – Costs (model calls, context handling, engineering effort) must be tied to business value. Wang Bo advocated early focus on user education and habit formation before aggressive cost‑cutting, while Wang Tonghuan suggested allocating stronger models and resources to high‑value scenarios and using cheaper paths for low‑value, batch workloads.

Future Architecture: Super Agent vs. Multi‑Agent Ecosystem – Wang Tonghuan predicts that standardised protocols will enable a powerful “Super Agent” that compresses many middle‑layer services, delivering data in near‑real‑time. The host, however, warned that a collaborative multi‑Agent ecosystem—each specialised in tasks like exploration, development, quality, security, or cost—may be more realistic for heterogeneous enterprise environments.

Conclusion – Transitioning to an Agentic data platform is not a simple AI overlay; it demands a paradigm shift in data infrastructure, governance, knowledge management, and ROI thinking. Companies that first build the necessary semantic, security, and feedback foundations will be able to cross the gap from demo feasibility to production reliability.