Agents as Database Clients: On the Incompatibility of Agentic AI Workloads with the ACID Transaction Contract

This preprint examines agentic AI systems as an emerging class of database clients and analyzes how their probabilistic, long-horizon, epistemic, and semantically dependent workloads stress the expressive scope of the classical ACID transaction contract. The paper argues that ACID remains valid for classical, explicitly bounded transactions, but that agentic workloads increasingly require additional database-level guarantees over longer semantic units of work. The manuscript identifies four structural mismatches between agentic workloads and classical transaction assumptions: epistemic rollback, semantic consistency, semantic conflict, and semantic expiry. It introduces three concepts, semantic consistency, semantic conflict, and semantic expiry as vocabulary for reasoning about database-managed state consumed and modified by autonomous agents. The paper proposes an initial research agenda toward an Agentic Database Contract, including long-horizon transaction semantics, semantic consistency models, intent-aware concurrency control, parameterized durability, and observability primitives for agentic workloads.