Databases have operated under an implicit contract for four decades: human-authored applications issue predictable queries, writes are intentional, connections remain brief. This foundation shaped schema design, connection pooling, and permission structures. AI agents now violate this fundamental assumption, issuing unpredictable queries based on reasoning paths rather than pre-written code, creating unprecedented challenges for traditional database architectures.

The most dangerous assumption is that writes were reviewed by humans. Agents write autonomously, potentially in loops or on retries, processing thousands of transactions before anyone notices. The solution requires designing write paths assuming the caller might be wrong. Soft deletes and append-only event logs provide safety nets, while idempotency keys prevent duplicate writes from retries, ensuring data integrity even when agent reasoning fails.

Connection models also face disruption as agents hold connections longer during reasoning steps and fan out to multiple simultaneous sessions. Traditional connection pools designed for brief human interactions break under these workloads. The fix requires dedicated connection pools sized specifically for agent workflows, with timeouts measured in seconds rather than minutes, addressing the fundamental mismatch between old assumptions and new AI-driven patterns.