Intelligence Becomes Free — What's Next? Data Systems for Agents, of Agents, and by Agents

Authors from the EPIC Data Lab at the University of California, Berkeley analyze the implications of declining costs for AI model inference. Over the past year, inference prices have dropped by a factor of 9 to 900, with a median reduction of about 50 times, making intelligence sufficient for most cognitive tasks nearly free. This gives rise to three directions for data systems. First, data systems for agents: agents perform a large volume of work, including speculative queries (agentic speculation), where up to 80-90% of sub-queries are redundant, offering opportunities for optimization—such as result reuse and multi-query optimization. Second, data systems from agents: infrastructure is needed to manage thousands of agents—organized memory (structured instead of unstructured Markdown files), coordination, and conflict resolution for concurrent changes. Third, data systems created by agents: agents can synthesize entire data systems on the fly, but their correctness must be verified. The authors discuss the concept of structured memory for agents, where memories are organized by attributes (e.g., module, language, operation), as well as issues of collaborative editing and livelock when multiple agents work simultaneously.