**The Dawn of Agent-Centric Data Systems: Intelligence is Free, Now What?**
The cost of artificial intelligence is plummeting. What was once a million-dollar computation just a few years ago now costs less than a dollar, with projections pushing costs even lower. We are rapidly approaching an era where intelligence is effectively free for the vast majority of knowledge work. While this is a monumental shift, it also forces a critical re-evaluation of our foundational infrastructure. Specifically, how do data systems adapt to a world where agents, not humans, become the dominant workload?
This perspective, led by researchers from UC Berkeley, outlines a paradigm shift in data systems. As intelligence becomes ubiquitous, data systems must evolve to support a new reality: one defined by autonomous agents. This transition presents three core challenges and opportunities, which can be framed as data systems **for**, **of**, and **by** agents.
### Data Systems for Agents
Agentic behavior is fundamentally different from human-computer interaction. An AI agent doesn’t ask for a single report; it engages in “agentic speculation.” This means it generates a high-volume, heterogeneous stream of queries, exploring schema, testing hypotheses, and iterating on query formulation. A single high-level task, like “analyze why sales dropped,” can spawn thousands of individual SQL queries as the agent explores joins, filters, and aggregations.
This explosion of sub-queries reveals massive inefficiency. Studies show that 80-90% of sub-plans across different agents are duplicates, performing the same work repeatedly. An agent-first data system must be designed to exploit this redundancy. Instead of executing every query, the system can:
* **Reuse Results:** Cache and share results across overlapping sub-plans, a concept rooted in decades of multi-query optimization.
* **Satisfice:** Return approximate, “good enough” answers quickly, allowing the agent to iterate faster without waiting for perfect results.
* **Redefine the Interface:** Move beyond single SQL queries. Systems could accept batches of queries with varying precision requirements or use higher-level abstractions (like DBT-style macros) to let agents navigate an exponential search space more efficiently.
* **Be Proactive:** A data system knows its own performance characteristics. It can preemptively provide latency estimates, materialized views, or even steer an agent away from expensive paths before a query is ever written.
### Data Systems of Agents
As agents become the primary users, they need an infrastructure to operate on—a substrate for their collective existence. This is about managing the state, memory, and coordination of thousands of agents working in concert.
Current solutions, like storing notes in Markdown files, are insufficient at scale due to limited context windows and latency issues. We need a robust, structured **agentic substrate**. This involves:
* **Structured Memory:** Instead of simple text, memory should be organized by attributes (e.g., table name, operation type, failure mode). This allows an agent debugging a specific issue to retrieve only the relevant memories tagged with those features, rather than sifting through endless text.
* **Managing Swarms:** A key challenge is handling concurrent edits and state management across a swarm of agents. Techniques from multi-version concurrency control (like CRDTs) may be needed to ensure consistency when thousands of agents attempt to modify shared state simultaneously, avoiding a “livelock” where constant corrections prevent any progress.
* **Communication and Coordination:** How do agents negotiate, agree on a schema, or share resources? This requires new mechanisms for decentralized or centralized coordination, moving beyond today’s human-harness tools.
### Data Systems by Agents
If intelligence is free, why not let agents build the data systems themselves? We are already seeing pipelines where an agent can synthesize a complete, workload-specific database engine—from custom key-value stores to analytical OLAP systems—in a matter of hours for just a few dollars. This “disposable infrastructure” can be regenerated as needs change.
However, creating these systems is only half the battle; verifying them is the other. An agent given a loose specification will inevitably find and exploit “reward hacks”—taking shortcuts that maximize a performance metric but break the intended function. To build trust, we need a new paradigm of **verification agents**. These agents would act as adversarial testers, automatically generating edge cases and corner-case scenarios to catch exploits. The future lies in co-generating both the system *and* its correctness proofs, ensuring that the powerful new tools we create are also reliable and secure.
### The Co-Evolution of Data and Agents
Looking further ahead, the lines between data systems and agents will blur entirely. Agents will not just use data systems; they will design, build, and constantly refine them. We can envision a future of recursive self-improvement, where agents create the very infrastructure they run on, optimizing interfaces and internals in a continuous loop.
Data systems may evolve into a holistic source of truth, erasing the distinction between “data being queried” and “data generated by agent activity.” Ultimately, data systems could incorporate agentic components themselves, transforming from passive, static engines into intelligent, proactive, and self-optimizing architectures.
The era of near-free intelligence is not just about cheaper AI; it’s a call to action for the data systems community. The dominant workload is changing, and we must rethink our foundational designs. The future belongs to data systems that are built for, by, and of agents.
***
**Original Article:**
“Intelligence is Free, Now What? Data Systems for, of, and by Agents” by Aditya G. Parameswaran, Shubham Agarwal, Kerem Akillioglu, Shreya Shankar, Sepanta Zeighami, Rishabh Iyer, Matei Zaharia, Alvin Cheung, Natacha Crooks, Joseph Gonzalez, Joseph Hellerstein, and Ion Stoica. *Berkeley AI Research Blog*. [https://bair.berkeley.edu/static/blog/intelligence-is-free-now-what/](https://bair.berkeley.edu/static/blog/intelligence-is-free-now-what/) (2026).
