Article body
Full article
Summary: Hengshi Data Agent is not a product that asks users to choose among capabilities. Data Q&A, modeling, and visualization creation work as a stack: they coexist and collaborate to complete the analytics path from data preparation and metric modeling to visual delivery. That is why they are delivered as one integrated offering. Drawing on Hengshi Data Agent’s multi-agent collaboration practice, this article explains how the three capabilities form a complete Agentic BI system and why capability stacking matters more than product selection.
1. Capability Stacking, Not Product Selection
A complete BI workflow spans a broad set of skills, from data preparation and metric modeling to visual delivery. Preparing data requires an understanding of database schemas, missing values, and outliers. That is data engineering work. Defining metrics requires an understanding of business definitions, calculation logic, and lineage. That is data governance work. Visualization requires chart selection, layout design, and interaction logic. That is data visualization and design work.
Putting all three skill sets into one Agent is difficult. Contexts interfere with each other, and an error in one stage can weaken the whole workflow. A common industry response is to split the capabilities and let users select what they need. Hengshi Data Agent takes the opposite approach. Its three capabilities work together as a stack that closes the loop.
Data Q&A relies on the metrics semantic layer built through modeling to keep definitions accurate. Visualization creation depends on the data prepared through data Q&A and presents metrics according to the modeling definitions. Modeling results must also reach business users through data querying and visualization. Each capability enables the other two. That interdependence is why Data Agent is delivered as an integrated capability set rather than separate products.
2. Three Capabilities, Three Stages of One Product
Hengshi Data Agent includes three specialized Agents. They are not three products that customers buy separately. They divide one analytics agent system into three cognitive stages: understanding data, defining logic, and presenting results.
Table 1. Overview of the three stacked capabilities
| Capability | Stage | Typical Output | Role in the Stack |
|---|---|---|---|
| Agent 01: Data Q&A Agent | Understand data | Standardized datasets | The translator for the data world, turning business requirements into accurate, safe data retrieval operations |
| Agent 02: Visualization Creation Agent | Present results | Dashboards and reports | The presentation specialist, turning analysis results into readable, useful dashboards |
| Agent 03: Modeling Agent | Define logic | Structured metric definitions | The architect of the metric system, turning business definitions into a unified semantic layer |
2.1 Agent 01: Data Q&A Agent, the Translator for the Data World
The Data Q&A Agent turns natural-language requirements from business users into data-layer instructions. It does not focus on how to write a piece of SQL. It identifies what data the user needs, where the data lives, and how to retrieve it.
When a user says, “I want to see the sales trend for the last six months,” the Agent does not write SQL immediately. It first identifies the metrics associated with sales, interprets the time meaning of the last six months, determines the relevant tables and fields, and then calls the tools that execute the query. The Data Q&A Agent does not produce the final analysis conclusion or visualization. It retrieves the correct data accurately and safely, then passes the result to the next stage.
2.2 Agent 03: Modeling Agent, the Architect of the Metric System
The Modeling Agent helps users build and maintain a metric system. Modeling here means BI data modeling, not machine-learning modeling. It defines metric calculation logic, dimensional hierarchies, and lineage between metrics.
For example, when defining customer lifetime value, the Modeling Agent guides the user through key questions: Does CLV mean historical cumulative spend or predicted value? What is the calculation period? Which tables and fields provide the data? How does the metric relate to other metrics? The Agent produces structured metric definitions rather than dashboards or reports. It writes those definitions to the metrics semantic layer, where they become the shared definition for future data querying, dashboards, and reports. The quality of the model determines how far the analysis above it can go.
2.3 Agent 02: Visualization Creation Agent, the Presentation Specialist
The Visualization Creation Agent presents analysis results in the most suitable visual form. Its core capabilities include recommending chart types from the data and analytical goal, using line charts for time series, bar charts for category comparisons, pie charts for proportions, and scatter plots for correlations. It can also design a dashboard layout automatically, placing metric cards at the top, trend charts in the middle, and detail tables at the bottom. Users can adjust chart styles and layouts through natural language.
Its distinct value is a lower barrier to visualization. Business users describe what they want to see, and the Agent handles how to present it.
3. Three Collaboration Patterns for the Capability Stack
When the three capabilities work as a stack, they collaborate in three typical patterns: a sequential pipeline, star scheduling, and federated collaboration.
Table 2. Comparison of the three collaboration patterns
| Collaboration Pattern | Typical Scenario | Lead and Supporting Roles |
|---|---|---|
| Sequential pipeline | Building a complete analytics dashboard from scratch, including data retrieval, definitions, and visualization | The Visualization Creation Agent initiates the flow, the Data Q&A Agent supports it, and the Modeling Agent provides definition assurance |
| Star scheduling | Building an enterprise-wide system of consistent metric definitions | The Modeling Agent leads, with support from Data Q&A and visualization creation |
| Federated collaboration | Workflows embedded in external Agent platforms such as Dify and Coze | The external platform orchestrates the workflow, while Hengshi Data Agent serves as the BI execution endpoint |
3.1 Sequential Pipeline: An End-to-End Path from Data to Dashboard
A marketing director asks for a Q2 sales performance dashboard for East China. The Visualization Creation Agent finds that it does not have a ready dataset, so it asks the Data Q&A Agent to prepare an East China Q2 sales dataset. The request includes sales revenue, order volume, and average order value, segmented by region and product category.
The Data Q&A Agent checks the data catalog, confirms the current user’s viewing permissions, retrieves and aggregates the data under the requested conditions, and returns a structured dataset to the Visualization Creation Agent. The Visualization Creation Agent plans the dashboard around the time-series and categorical data, renders an initial version, and refines it with the user. The user made one request. The Modeling Agent did not enter the exchange directly, yet its existing definitions for sales revenue and average order value supplied the definition assurance for the full process. Each stage has the right capability available.
3.2 Star Scheduling: Metric System Building Led by the Modeling Agent
A retail enterprise CIO launches an AI-Ready Data initiative to establish consistent metric definitions on the Hengshi platform. The Modeling Agent leads the work. It first collaborates with the Data Q&A Agent to inspect schemas and understand the available data sources, table fields, and existing metrics. It then works with business owners to define core metrics one by one and checks their dependencies automatically.
For example, if average order value is defined as sales revenue divided by order count, the Agent automatically links the existing definitions of those two metrics. After the definitions are complete, the Visualization Creation Agent creates a metric overview dashboard for business acceptance and confirmation. This pattern follows the goal of establishing a metric system. The Modeling Agent chooses the collaborators and the way they work together based on the task at hand.
3.3 Federated Collaboration: Working with External Agent Platforms Such as Dify and Coze
Together, Hengshi Data Agent and HENGSHI CLI function as a BI execution endpoint. They can join a broader Agent ecosystem as tool nodes in external platforms such as Dify and Coze. The external platform understands high-level intent and orchestrates the business workflow. The Hengshi side exposes standard tools for querying data, creating dashboards, and defining metrics.
Consider an intelligent operations analytics workflow. Dify triggers the workflow on a schedule and calls Hengshi’s data Q&A tools to retrieve previous-day sales, traffic, and conversion data across all channels. Dify passes the data to an internal anomaly-detection module. If the module finds an abnormal metric, Dify calls Hengshi Data Agent again for a multidimensional drill-down and root-cause analysis, then formats the result as a daily report. Hengshi Data Agent acts as a specialist node, focusing on BI analysis. The external platform can orchestrate each capability as a node while the product remains one integrated offering.
4. Mechanisms That Make the Stack Work
4.1 A Unified Metrics Semantic Layer: A Common Language Between Agents
The metrics semantic layer is the key infrastructure that lets the three capabilities collaborate without friction. The Data Q&A Agent labels retrieved data with metric definitions. The Modeling Agent persists its metric definitions in the semantic layer. The Visualization Creation Agent automatically associates the charts it displays with those definitions. The three Agents share the language of metrics instead of having to interpret each other’s specialized language.
4.2 Shared Analysis State: The Handoff Record Between Agents
Agents transfer more than data. They also transfer analysis state: the current business domain, the filters already applied, and the user’s priorities. This metadata moves with the handoff. It resembles a hospital registration record. A physician in the next department can see the tests and preliminary diagnosis already completed without asking the patient to start again.
4.3 Fault Isolation and Graceful Degradation
If the Visualization Creation Agent has a performance issue, the Data Q&A Agent can continue serving queries. If the Modeling Agent is undergoing maintenance, users can still query existing metrics through the Data Q&A Agent. An issue in one Agent does not take down the entire system, which matters in enterprise applications.
5. Frequently Asked Questions
Q1: Does multi-Agent collaboration add latency?
Multi-Agent collaboration does add internal communication beyond a simple, single-step Agent response, but those exchanges are usually measured in milliseconds. Data query time remains the main driver of total latency. Each Agent also focuses on a narrower domain with a more concise context, so its own reasoning can be faster.
Q2: Can an enterprise deploy only one of the three Agents?
No. The three capabilities work as a stack and need to exist together to perform well. Data Q&A depends on the metrics semantic layer produced through modeling. Visualization creation depends on ready data and consistent metric definitions. A single capability cannot realize its full value without support from the other stages. That is why Data Agent is not sold as separate capabilities.
Q3: Can multiple Hengshi capabilities be called as independent nodes in a Dify or Coze workflow?
Yes. As capability nodes of the same product, Data Q&A, modeling, and visualization creation can all register with platforms such as Dify. You can orchestrate them around your business logic, for example by calling Data Q&A in the first step and visualization creation in the second. The product is delivered as a whole, while its capabilities remain available for external orchestration.
Q4: Which Agent initiates collaboration among the three?
The task determines the lead. In a pipeline flow such as retrieving data and then creating charts, the Visualization Creation Agent initiates scheduling. In exploratory analysis such as metric modeling with validation, the Modeling Agent leads. In direct data Q&A, the Data Q&A Agent responds and decides whether to involve other Agents. No fixed primary Agent exists. The Agent best suited to the current task schedules the work.
6. Conclusion
Multi-Agent collaboration responds to a practical fact: BI workflows span too many stages and too many skills for one generic Agent to cover well. Hengshi Data Agent Family shows that three specialized Agents can work together as a capability stack more reliably than asking users to choose between them. In Agent design, assembling the right stack matters more than making an isolated choice.