Agentic systems are usually discussed in terms of models and prompts. This project reads them the other way round, as data systems: what they store, what they retrieve, what they pass between each other, and what breaks when any of it is wrong.
The survey, Data in Agentic AI: A Comprehensive Survey, organises the literature into nine parts:
- Data processing — data agents, construction, management and selection
- Agentic RAG and GraphRAG — retrieval-augmented generation over graph structure
- Memory systems — long-term memory, context management, KV efficiency
- Multi-agent systems — orchestration, collaboration, planning, communication
- Self-evolving agents — reflection, learning, self-improvement
- Security, privacy and safety — prompt injection, privacy leakage, trustworthiness
- Benchmarks, tools and protocols — MCP, tool use, evaluation frameworks
- Applications — coding, finance, scientific discovery, search
- Related surveys
The companion tutorial, Data-Centric Foundations of Agentic AI, was presented at ICDE 2026.
The resource list is maintained alongside the paper and is open to contributions.
Authors. Yuxin Jin, Hanchen Wang, Ying Zhang, Dong Wen, Lu Qin, Wenjie Zhang.