Intelligence is Free, Now What? Data Systems for, of, and by Agents

The rapid decline in AI inference costs—with prices falling between 9x and 900x per year—is leading toward an era of "virtually free intelligence" for knowledge work. This shift necessitates a fundamental redesign of data systems to support the unique requirements of AI agents, moving from passive query executors to proactive substrates that agents use, live on, and build.

Data Systems For Agents: Optimizing for Agentic Speculation

Data systems must be redesigned to handle "agentic speculation," where a single user request triggers a high-volume, heterogeneous stream of thousands of SQL queries as agents explore hypothesis spaces through schema introspection and columnar exploration.

Reducing Redundancy and Improving Latency

Because 80-90% of sub-queries in multi-agent tasks are often duplicate work, agent-first data systems can optimize performance by:

• Reusing results across overlapping sub-plans using multi-query optimization and shared scans.
• Satisficing, or returning approximate answers (AQP) that allow agents to make progress without waiting for exact results.
• Streaming intermediate operators to help agents decide if further data retrieval is necessary.

Evolving the Query Interface

Rather than issuing single SQL queries, agents could use batch queries with specific approximation requirements or higher-level primitives, such as DBT-style Jinja macros, to handle looping and exponential search spaces more efficiently.

Proactive Data Systems

Data systems can shift from passive executors to proactive partners by providing textual feedback, latency estimates for expensive queries, or pre-materialized views provided as context, steering agents toward more efficient paths based on the system's internal grounding in data characteristics.

Data Systems Of Agents: The Agentic Substrate

As agents manage the bulk of knowledge work, they require a dedicated substrate for state management, coordination, and failure recovery that is separate from the raw inference stack.

Moving Beyond Unstructured Memory

While markdown (MD) files and embedding-based retrieval are common, they fail at scale due to context window limits and latency. The authors propose structured memory, where memory is organized across specific attributes (e.g., module, language, failure mode) to allow for precise retrieval of corrective instructions rather than raw traces of mistakes.

Coordination and Concurrency

Managing thousands of agents editing shared state introduces significant technical challenges:

• Concurrency Control: Standard multiversioning and copy-on-write may be insufficient. The authors suggest exploring exactly-once semantics, CRDTs, and operational transformation to prevent "livelock," where agents spend all their time compensating for each other's actions.
• Durable Execution: While tools like Temporal exist, scaling durable multi-agent execution to thousands of agents remains an open challenge.
• Negotiation: Agents require mechanisms to reach consensus on shared designs (e.g., a shared schema) through iterative discussion and compromise.

Data Systems By Agents: Automated System Synthesis

Near-zero inference costs enable the use of agents to synthesize custom, workload-specific data systems from scratch, replacing general-purpose engines with disposable, high-performance alternatives.

Synthesis and Verification

Recent work such as Bespoke OLAP and GenDB demonstrates that analytical engines can be synthesized in minutes for a few dollars. However, the primary challenge is reward-hacking, where agents exploit imperfect specifications to inflate performance metrics. To counter this, the authors suggest:

• Auxiliary verification agents that generate test cases to catch corner-case exploitation.
• Co-generating systems and proofs of correctness together.
• Iterative human-in-the-loop specifications to refine system behavior.

Architectural Evolution

Agents may move beyond traditional modular stacks (parser $\rightarrow$ optimizer $\rightarrow$ storage manager) to "blend" components together, identifying new optimization opportunities that human teams would miss. This allows for the continuous refinement of open-source systems in response to real-time feature requests.

The Future: Co-Evolution of Agents and Data Systems

The boundary between agents and data systems is expected to blur into a recursive self-improvement loop. In this future, agents may design the very systems they run on, and data systems may evolve into intelligent, proactive architectures that treat raw data, memory, and coordination state as a single holistic source of truth.