Sorting Data Agents Into Six Levels of Autonomy—and Hype

What it does The repository is a curated, continuously updated paper list that sorts research on LLM-powered data agents into a six-level autonomy taxonomy modeled on the SAE J3016 standard for driving automation. It indexes work across data management, preparation, and analysis, from L0 (manual) through L5 (theoretical full autonomy), and serves as the companion bibliography to the survey A Survey of Data Agents: Emerging Paradigm or Overstated Hype? and a SIGMOD 2026 tutorial.

The interesting bit Rather than treating “agent” as a binary label, the taxonomy explicitly buckets most existing systems as L1 assistants or L2 procedural executors, while relegating L4–L5 to a “vision” section. That framing makes the repo unusually honest: it functions as a structured reality check on a field the authors warn suffers from terminological ambiguity and mismatched expectations.

Key highlights

• Six-level autonomy scale (L0–L5) that traces progressive shifts in human-to-agent responsibility.
• Papers indexed by tier and task: Data Management, Data Preparation, and Data Analysis.
• Companion to an arXiv survey and a 3-hour SIGMOD 2026 tutorial on data-agent levels and open problems.
• Includes a KDD Cup 2026 track on Data Agents for Complex Data Analysis.
• Community-driven: accepts new paper recommendations via GitHub issues.

Caveats

• This is a reading list and taxonomy document, not a runnable codebase or framework.
• The upper tiers (L4–L5) are explicitly forward-looking and lack corresponding implemented systems in the list.

Verdict A useful bookmark for researchers and builders who need to navigate the LLM data landscape without drowning in vague “agent” branding. Not a source of ready-to-use software.