Awesome Data Agents — What This Project Is

Awesome Data Agents is a curated reading list (a collection of research papers and resources) about data agents — software systems that use large language models (LLMs) to help with data-related tasks like managing databases, cleaning data, and analyzing information.

The Core Idea

The project's authors noticed that the term "data agent" is used loosely — some people use it for simple chatbots that answer questions, others for fully autonomous systems. To fix this confusion, they wrote a survey paper (a review of existing research) that proposes a six-level scale (L0 to L5) for classifying how much autonomy a data agent has:

• L0 — No automation; humans do everything manually.
• L1 — The agent assists with single, simple tasks (e.g., generating code snippets) but needs heavy human checking.
• L2 — The agent can perceive its environment (e.g., read data lakes, call APIs) and execute procedures, but humans still orchestrate the overall workflow.
• L3 — The agent becomes the dominant actor, with a human only supervising.
• L4 — High autonomy; the agent acts proactively with minimal human oversight.
• L5 — Full autonomy; the agent is fully generative and self-directed (a future vision, not yet realized).

The survey argues that moving between levels requires major capability leaps — not just incremental improvements.

What the Repository Contains

• A paper list organized by autonomy level (L0–L1, L2, proto-L3, and L4–L5 prospects), and within each level, by task type: Data Management (e.g., database tuning, query optimization, system diagnosis), Data Preparation (e.g., cleaning, integration, discovery), and Data Analysis (e.g., table question-answering, text-to-SQL, visualization, report generation).
• Comparison tables for agents at each level, noting techniques used (e.g., in-context learning, retrieval-augmented generation, fine-tuning) and data complexity handled (multi-source, heterogeneous, multimodal).
• Related resources: benchmarks for data agents, plus other surveys and tutorials.
• Research opportunities section (listed in the table of contents).

Notable Facts

• The survey paper is available on arXiv (October 2025).
• The authors presented a 3-hour tutorial at SIGMOD 2026 (a major database conference) — slides are included in the repo.
• The project launched a KDD Cup 2026 competition track on "Data Agents for Complex Data Analysis" (official site: dataagent.top).
• The repo invites community contributions: you can recommend a paper by opening a GitHub issue, and contact the authors via email.

How to Use It

If you're a researcher or practitioner working on LLM-powered data tools, this repo serves as a structured map of the field — helping you see where existing work sits on the autonomy spectrum, what's been done, and where the gaps are (especially the L2-to-L3 transition, which the authors highlight as a critical open problem).