Supporting Our AI Overlords: Redesigning Data Systems to Be Agent-First (arxiv.org)

A new paper argues that LLM-driven agents—models acting autonomously to query, manipulate and analyze data—are poised to become the dominant workload for data systems, and that current databases and storage stacks are not built for how those agents operate. The authors coin the term “agentic speculation” to describe agents’ high‑throughput pattern of exploratory queries and iterative solution refinement. That behavior generates extreme scale, heterogeneous access patterns, heavy redundancy, and strong steerability (agents being guided by feedback), creating performance, cost, and correctness challenges for existing systems. The paper outlines an “agent‑first” data systems agenda: rethink query interfaces to express intent, uncertainty, provenance and agent state; design query processors that support massive speculative execution, deduplication/pruning of redundant work, incremental and merged result materialization, and cross‑agent coordination; and build new agentic memory stores optimized for recall, long‑term state, and fast retrieval of previous agent interactions. These directions have concrete implications for indexing, caching, concurrency control, cost accounting, and privacy. For the AI/ML community this signals research and engineering opportunities at the intersection of databases and agents—new APIs, algorithms and architectures to make agent workloads efficient, auditable and scalable.