Article body
Full article
The first experience with conversational analytics is compelling: ask a question and get data back. A few weeks later, most teams hit the same wall. Analysts still maintain datasets, build charts, watch for anomalies, and transfer conclusions into another business system. A chat interface lowers the query barrier, but it does not shorten the full chain of work.
The value of a Data Agent depends on how much real work it can complete. HENGSHI SENSE 6.2, for example, separates modeling, creation, querying, and page operations into distinct capabilities. A user can ask the system to join order and customer data, build a regional sales dashboard, and explain a decline in East China. A task planner breaks the goal into steps, then coordinates semantic modeling, metric retrieval, chart configuration, and explanation. The result is a workflow with visible progress, intermediate evidence, and a concrete deliverable.
From a Business Goal to a Task Graph
A compound request needs planning before execution. The system must identify data objects, business metrics, time ranges, and the expected deliverable. It must also make dependencies explicit. A dataset must be confirmed before a dashboard is created. A standard calendar definition must be selected before year-over-year growth is calculated. Query results must be validated before an anomaly is explained.
The task graph makes those dependencies reviewable and lets a user pause, revise, or rerun one step. A business user may only say, “Find out why East China sales dropped.” The agent has to translate that goal into metric confirmation, trend analysis, regional and product breakdowns, and anomaly attribution, retaining the input, tool call, and result at each step.
Four Capabilities, One Workflow
| Capability | Primary responsibility | Typical deliverable |
|---|---|---|
| Modeling agent | Discover data, validate fields, configure joins and metrics | Reusable dataset or semantic model |
| Creation agent | Map analytical intent to charts and page configuration | Dashboard, chart, or report |
| Query agent | Plan and run queries, then explain the result | Data, evidence, and conclusion |
| Navigation agent | Navigate, change configuration, and explain the action | Visible page change and execution receipt |
These are not four isolated chatbots. The task planner owns dependencies, timeouts, retries, and human takeover. Shared state lets a modeling output become the input for subsequent creation and query steps.
The Semantic Layer Defines the Answer Boundary
A language model can understand words, but it cannot infer whether a company’s revenue metric includes tax or whether an active customer means a login in seven days or a payment in thirty. The Data Agent must retrieve metric definitions, dimensions, permissions, and lineage from the semantic layer before it generates a query or configuration. Shared metrics keep dashboards, conversational queries, and APIs consistent.
The semantic layer also resolves ambiguity. When a user asks for “sales,” the system can show candidate metrics with definitions and scopes. One brief confirmation is cheaper than allowing a report with the wrong business definition into an operating review.
Headless APIs Make Execution Governable
A reliable agent does not automate mouse coordinates. It uses Headless APIs to create datasets, configure joins, build charts, and change styles. APIs can validate parameters, record audit events, and return structured errors. A page assistant remains useful for navigation and explanation; governed interfaces should handle changes to core resources.
Read and write operations need different controls:
- Queries can run within the current user’s data permissions.
- Resource creation starts with a preview of targets, dependencies, and impact.
- Changes to critical metrics or sharing scope require approval.
- Cross-system actions need idempotency, rollback, or compensation.
- Agent permissions never exceed the permissions of the current user.
Broad Context Does Not Mean Broad Authority
A Data Agent may cover many product areas, but every call remains bound to the current tenant, user, and resource policy.
| Area | Covered capability | Example operation |
|---|---|---|
| Application creation | Dashboard and page management | Create a sales analytics dashboard |
| Data marketplace | Package browsing and resource discovery | Locate an East China sales dataset |
| Dashboard | Chart operations and interactions | Change a bar chart to a line chart |
| Dataset | Field and relationship management | Add a customer dimension |
| Data connection | Source management and connection tests | Check a MySQL connection |
| Data pipeline | ETL workflow configuration | Schedule a daily synchronization |
| Permissions | Role and data authorization | Grant read-only regional access |
| API management | Key management and usage statistics | Review API call volume |
| User management | Accounts and organization structure | Create a data analytics team |
Start with One Short, Measurable Chain
A first deployment should target a frequent task with clear boundaries, such as checking core metrics every morning and producing an anomaly summary, or generating a weekly dashboard from an approved dataset. Measure task success, metric clarification, human takeover, tool retries, and audit completeness before expanding into modeling or cross-system execution.
A mature Data Agent is not defined by how human its conversation sounds. It is defined by how reliably it completes a task and how clearly a user can inspect what it did.
Further reading: HENGSHI SENSE Product and Technology White Paper.