Briefing Document: A Hierarchical Taxonomy for Data Agents (L0-L5)

This document summarizes the systematic hierarchical taxonomy proposed for classifying Data Agents. A Data Agent is defined as a comprehensive, Large Language Model (LLM)-powered architecture that orchestrates the Data + AI ecosystem to autonomously perform a wide range of data-related tasks. These tasks span the data lifecycle, including Data Management (Configuration Tuning, Query Optimization, System Diagnosis), Data Preparation (Cleaning, Integration, Discovery), and Data Analysis (Structured/Unstructured Analysis, Report Generation).

The term "data agent" currently suffers from terminological ambiguity and inconsistent adoption, conflating simple query responders with sophisticated autonomous architectures. This ambiguity leads to: • User-Side Risk: Expectation mismatches regarding the agent's scope and limitations, which can lead to undue reliance on erroneous outputs. • Governance Risk: Indistinct lines of responsibility when failures occur (e.g., data leakage or erroneous reports), challenging accountability. • Industry-Side Risk: Erosion of market confidence due to overstated claims and blurred capabilities, hindering objective system comparison. Inspired by the SAE J3016 standard for driving automation, this taxonomy introduces six levels (L0-L5) that delineate and trace progressive shifts in autonomy, clarifying capability boundaries and responsibility allocation. *

The hierarchy is structured around the progressive transfer of dominance and responsibility from the human to the data agent. | Level | Autonomy Designation | Human Role (Control & Responsibility) | Data Agent Role (Functionality) | Evolutionary Leap | | :---: | :---: | :--- | :--- | :--- | | L0 | No Autonomy | Human in charge; All tasks are entirely human-driven. Humans are solo practitioners. | Uninvolved. | N/A | | L1 | Assistance | Humans retain task dominance and responsibility for integration, verification, and interacting with the environment. Humans shift from solo practitioners to users of query-responsive assistants. | Provides preliminary, stateless assistance for isolated tasks. Functions as a nascent intelligent assistant within a prompt-response framework. | Introducing Assistance (L0 to L1) | | L2 | Partial Autonomy | Humans are responsible for managing the overall workflow and retain dominance over tasks (e.g., orchestrating pipelines). | Gains the ability to perceive and interact with the environment (e.g., data lakes, tools, APIs). Acts as a procedural executor within human-orchestrated pipelines. | Gaining Perception (L1 to L2) | | L3 | Conditional Autonomy | Humans act as supervisors overseeing the agent’s operation, but the agent assumes the dominant role in tasks. | Autonomously orchestrates and optimizes tailored data pipelines for diverse and comprehensive data-related tasks. Evolves from procedural executor to a versatile dominator. | Transfer of Task Dominance (L2 to L3) | | L4 | High Autonomy | Humans fully delegate responsibility, becoming onlookers and recipients of insights. | Achieves high autonomy and reliability, operating independently of human supervision. Can proactively identify issues and autonomously orchestrate pipelines to tackle self-discovered problems. | Removing Supervision (L3 to L4) | | L5 | Full Autonomy | Complete disengagement; any form of human involvement is unnecessary. | Functions as a generative data scientist capable of inventing novel solutions and pioneering new paradigms. | Innovating and Pioneering (L4 to L5) | *

The field is currently focused on the L2-to-L3 transition, where agents must evolve from procedural execution to autonomous orchestration.