AI Agents May Be the Workload DBOS Was Waiting For

AI agents may present the workload DBOS was seeking. While DBOS initially proposed representing operating system state as database tables, a more practical approach may be to view the database as the durable runtime for autonomous computation. Agents, unlike traditional web requests, have a long-running execution state that requires durability beyond simple retry logic.

They accumulate various state stores, including Postgres for business state, Redis for caching, vector databases for semantic retrieval, workflow databases for execution history, object storage for files, and tracing databases for model and tool traces.

This workload is well-suited for representation as a transactional database. Agent messages, tool-call results, approvals, checkpoints, permissions, budgets, execution history, and references to artifacts all constitute structured state that can be stored in a database. An agent runtime can be conceptualized as processes, threads, schedulers, files, system calls, credentials, process trees, resource limits, checkpoints, and provenance, all of which can have a durable database representation.

The 2025 DBOS retrospective highlights the importance of provenance, particularly when the program itself is making decisions. Queries over execution state, such as understanding why an agent took a particular action, which model response caused it, what data it saw, which tool it invoked, and how much the path cost, are naturally expressed through database queries. This presents a strong case for DBOS as the durable runtime for AI agents.