What Database Researchers Can Contribute to Agentic AI

Agentic AI systems are often described as orchestration layers that stitch together models and tools. That framing is useful, but it leaves out the systems perspective that database researchers bring.

The first contribution is planning. A useful agent does not merely retrieve information; it chooses among resources with different costs, constraints, and semantics. Databases already model this problem through query planning, cost estimation, and execution strategies. Those ideas transfer directly to agentic systems where tools, APIs, models, and datasets all compete for attention.

The second contribution is provenance. An answer or action produced by an agent is only as trustworthy as the sequence of steps that produced it. Database systems have long treated lineage, transactions, and execution traces as first-class concerns. Agent systems need the same discipline if they are going to support scientific or operational use cases.

The third contribution is reliability under mixed state. Many real-world tasks involve temporary state, cached state, external side effects, and partial failures. Database systems already reason about isolation, recovery, idempotence, and repeated execution. Agent stacks need comparable guarantees when they move beyond demos and into settings where errors matter.

The fourth contribution is heterogeneity. Scientific and operational workflows rarely stay inside one engine. Query processors and polystores offer a vocabulary for spanning relational data, graph data, search, files, and API-backed services without pretending that everything is a table.

Database researchers can also help with evaluation. Agent benchmarks often focus on answer quality and overlook the systems properties that determine whether a workflow is usable at scale. Latency, plan stability, provenance, recovery behavior, and observability should matter just as much as single-turn success rates.

References

P2KG: Declarative Construction and Quality Evaluation of Knowledge Graphs from Polystores. PROFORMA: Proactive Forensics with Message Analytics. Generating Polystore Ingestion Plans: A Demonstration with the AWESOME System.

Related projects

• AWESOME Polystore
• National Data Platform Search

Related publications

P2KG, PROFORMA, and the AWESOME polystore papers.

I am interested in exchanging ideas with researchers and engineering teams working on related systems.