… authorities of the individuals, by the individuals, for the individuals …
— Abraham Lincoln, Gettysburg Tackle (1863)
The price of AI is dropping quickly. GPT-4-class capabilities value roughly $30 per million tokens in early 2023; right now the identical runs below $1, and some providers are pushing costs below $0.10. Throughout benchmarks, inference prices have fallen between 9x and 900x per year, with a median decline close to 50x. Even frontier models are getting dramatically cheaper every technology, with open-source fashions following carefully behind. And crucially, even when “Nobel-Prize-winning genius-level” intelligence isn’t right here but, the intelligence that suffices for the overwhelming majority of data work is right here right now, and getting cheaper by the month. At this price, we’re quickly getting into the period of just about free intelligence—the type that’s greater than sufficient for on a regular basis information work.
Disclosure: This publish is a perspective led by Aditya G. Parameswaran—an Affiliate Professor of EECS and co-director of the EPIC Knowledge Lab at UC Berkeley—collectively along with his collaborators. It’s half panorama survey and half perspective, and a number of other of the analysis instructions mentioned under (together with agentic hypothesis, structured reminiscence, and synthesizing customized knowledge programs from scratch) draw on the authors’ personal ongoing work.
So, what does this new period of near-free intelligence imply for knowledge programs? We consider three new challenges—and alternatives—stem from near-zero inference prices:
Knowledge Methods For Brokers. Brokers will quickly turn into the dominant workload for knowledge programs—with swarms of brokers spun up in response to every end-user request. Given variations in traits between brokers and people—or purposes appearing on their behalf—how ought to we redesign knowledge programs for such agentic customers?
Knowledge Methods Of Brokers. As brokers begin taking over the majority of data work, a brand new substrate is required for 1000’s of brokers to handle state over long-running duties, coordinate and attain consensus, and take care of failures. What do knowledge programs that reliably and effectively run and handle agent swarms appear to be?
Knowledge Methods By Brokers. Brokers are quickly turning into able to synthesizing whole knowledge programs in a single go—which means we will rebuild customized programs for every new workload. Verifying that such programs match supposed conduct is a problem. What does it take to let brokers synthesize knowledge programs we will really belief?
Knowledge Methods For, Of, and By Brokers
Subsequent, we’ll focus on every in additional element, adopted by discussing the intertwined future of information programs and brokers, particularly because the three challenges intersect.
Knowledge Methods For Brokers
An agent querying a database doesn’t behave like an individual or a BI instrument. It performs what we name agentic speculation: a high-volume, heterogeneous stream of labor spanning schema introspection, columnar exploration, partial after which full question formulation. With a number of brokers every exploring parts of the speculation house, every consumer request may quantity to 1000s of particular person SQL queries. Now, customers can challenge ‘high-level’ knowledge duties, e.g., root-cause evaluation—e.g., ‘why did espresso gross sales in Berkeley drop this 12 months’—or exploratory cohort evaluation—e.g., ‘which consumer segments are most probably to churn subsequent quarter’—every involving a combinatorial house of potential joins, aggregations, and filter mixtures.
Knowledge Methods Redesigned to Extra Successfully Help Agentic Hypothesis
The requests from these brokers have varied alternatives for optimization. As an example, on a text-to-SQL benchmark with a number of brokers trying every activity, solely 10-20% of the sub-plans are distinct. Thus, 80-90% of sub-queries carry out duplicate work. The identical experiments present activity success charges considerably rising with extra agentic makes an attempt—so the redundancy is definitely useful. However from the information system perspective it’s wasted work.
An agent-first knowledge system can exploit such properties to assist brokers make progress sooner. It could possibly reuse outcomes throughout overlapping sub-plans, drawing on concepts from decades-old literature on multi-query optimization and shared scans. Or the information system can attempt to satisfice, returning approximate solutions which are adequate for brokers to make progress, leveraging work from the AQP literature—or streaming the outcomes of the ultimate or intermediate operators to assist brokers determine if seeing the remainder is important or useful.
One other alternative right here is to rethink the question interface solely: as a substitute of brokers issuing a single SQL question at a time, they may as a substitute challenge a batch of queries, every with its personal approximation necessities. Since enumerating an exponential search house (as within the root trigger or cohort evaluation examples above) isn’t use of agentic reasoning capability, maybe knowledge programs ought to assist higher-level primitives quite than requiring brokers to checklist every SQL question explicitly. One thought right here is to attract on DBT-style Jinja macros to offer looping-based primitives for brokers to work together with knowledge programs.
A Caffeinated Military of Brokers Able to Tirelessly Full Your Knowledge Duties
A remaining alternative right here is to cease considering of information programs as passive executors of queries; knowledge programs could possibly be proactive, as they possess extra grounding in knowledge and system traits that brokers could lack a priori—they may steer brokers in several instructions, present outcomes for associated queries, and likewise present performance-level suggestions (e.g., as a substitute of executing an costly question, the system may first present the agent a latency estimate). The rationale we will do that now versus the previous is that an agent can settle for any type of textual suggestions and isn’t anticipating a strict SQL question outcome. The truth is, the information system may additionally put together each materialized and digital views for an agent prematurely, supplied to the agent as a part of context, as this can be cheaper or more practical than having an agent writer or use them.
Knowledge Methods Of Brokers
Beforehand, we centered on how brokers work together with knowledge programs. Now, we contemplate all the things else brokers must hold working: the place they dwell, how they keep in mind, how they coordinate with one another, and the way they take care of failures of one another. This agentic substrate is separate from the inference stack powering uncooked intelligence. Nonetheless, the inference stack itself is being abstracted away by way of APIs (e.g., from OpenAI or Anthropic), or, for open-weight fashions, by way of serving frameworks that cover low-level particulars. Thus far, the agentic substrate has been managed by way of harnesses like Claude Code and Codex, coupled with varied mechanisms to store and retrieve reminiscence.
First, on the reminiscence entrance, the present knowledge is that files are all you need; brokers write to unstructured markdown (MD) information, which may then be searched utilizing grep, or by way of embedding-based retrieval. The truth is, many argue that the answer to continuous studying is having brokers eat quite a bit (e.g., a complete codebase, slack, firm wikis, …) after which write their learnings into MD information, that are then retrieved selectively on demand. Certainly, file programs, bash scripting, and MD information are and can nonetheless be vital for brokers. Nonetheless, at scale, when brokers are doing the overwhelming majority of data work, this method will not be efficient.
Given restricted context home windows, retrieving all MD file fragments that could be related and stuffing it into the context will break down sooner or later. Even when context home windows proceed to develop, there are latency advantages to not put all data into context — and in lots of instances, e.g., when information work includes interacting with giant databases or code bases, will probably be infeasible to serialize all related knowledge into context.
Knowledge Methods As A Substrate for Multi-Agent Swarms
One may use a knowledge graph representation, however information graphs endure from the identical limitations as unstructured MD-based reminiscence resulting from their lack of structured search. What one wants is to have the ability to retrieve solely reminiscence that’s pertinent to the duty, throughout a number of attributes (or sides) of curiosity. For instance, an agent debugging a flaky check ought to be capable to pull solely the reminiscences tagged with the related module, language, framework, and failure mode—quite retrieving based mostly on key phrases or embedding similarity. A separate challenge is what to truly retrieve; uncooked agent traces with errors are usually not very helpful as they are going to induce brokers to repeat the identical mistake—as a substitute, we wish the retrieved reminiscence to be corrective.
We just lately explored a associated notion of structured memory, the place we manage reminiscence throughout varied attributes, every of which could possibly be set as * to point common applicability, or set as an inventory of values to be matched. For an information agent, the scale may embrace the columns and tables, sort of operation, and eventually, open-ended natural-language corrective directions. So, we may embrace reminiscence that solely applies to a given sort of operation (e.g., ‘when performing date-time operations, use fiscal 12 months versus calendar 12 months conventions’), or a given desk (e.g., ‘column product_cleaned is most well-liked over column product when querying on product identify’). One open query is defining an application-specific structured reminiscence—or what others have referred to as world models for memory. We consider that is akin to defining a schema for every utility—and maybe brokers themselves may also help us outline and refine it over time.
One Doable Means To Retailer and Retrieve Structured Information [From Here]
Structured reminiscence might be helpful additionally for evolutionary frameworks to successfully handle search areas. Certainly, storing, structuring, and mining giant volumes of single and multi-agent traces may also help future brokers turn into rather more environment friendly—doubtlessly enabling efficient recursive self-improvement by way of structured memory-based mechanisms.
One other problem is to assist concurrent edits to shared reminiscence, and concurrent edits basically, when there are various brokers performing transformations. Whereas there have been some helpful makes an attempt at supporting multiversioning and copy-on-write semantics, it isn’t clear that such strategies will suffice when 1000’s of brokers try to edit shared state on the similar time. As an example, when brokers try varied potential transactions in response to a consumer request, the results of the overwhelming majority of those transactions must be rolled again—with solely the one ‘appropriate’ transaction’s outcome persisting. Work on supporting exactly-once semantics is related right here, as are underlying strategies based mostly on CRDTs and operational transformation. For updates to fuzzy mechanisms corresponding to reminiscence, we might be able to sacrifice on consistency for excellent correctness within the curiosity of latency. Whereas brokers can motive about semantics to compensate or roll again their actions to ultimately finalize most duties, the first problem lies within the diploma to which they step on one another’s toes through the course of. An vital failure mode to be averted is a type of “livelock,” the place incessant compensating actions stop any significant progress.
Past shared state, different issues emerge when attempting to assist a military of brokers, together with what to do when brokers fail, how brokers ought to talk with one another (immediately or by way of intermediate shared state), and the way we must always take care of straggler brokers. There have been some developments in supporting sturdy multi-agent execution, corresponding to Temporal, but it surely stays to be seen if such options will apply at scale throughout 1000’s of brokers. On the subject of communication, we’d like mechanisms to allow brokers to barter with one another. Think about 4 developer brokers trying to achieve consensus on a shared schema, with distinct however overlapping goals. In a human setting, this is able to contain iterative dialogue and compromise; for agentic swarms, we should outline the mechanisms that permit them to converge on a design that displays the underlying targets of their respective principals. Or if brokers are all requiring entry to a restricted useful resource, once more communication might be obligatory. It stays to be seen if that is finest achieved by way of centralized coordination, or if a decentralized method is important.
Knowledge Methods By Brokers
Lastly, if intelligence is successfully free, then we will make use of this intelligence to synthesize new knowledge programs from scratch. Certainly, in lots of settings, general-purpose knowledge programs could also be overkill, as they must assist each schema, question, and {hardware} goal. Given a workload, current work, together with Bespoke OLAP and GenDB, has proven that one can use an agentic pipeline to synthesize a whole, workload-specific analytical engine—in minutes to a couple hours, at a value of some {dollars}. The engines are disposable: when the workload shifts, one can merely regenerate them. Analogously, our work has proven that one can synthesize customized key-value stores from scratch, focused to the workload. The truth is, fashionable IDEs, corresponding to Kiro, elevate specs for programs improvement to be a first-class citizen.
Brokers Can Synthesize Customized Knowledge Methods From Scratch
The primary challenge, nonetheless, is that specs are sometimes imperfect, and don’t cowl all nook instances. Current-day brokers will exploit the lacking specs to reward-hack their technique to a excessive efficiency metric. In our customized key-value retailer work, we discovered that one technique to alleviate that is to have auxiliary verification brokers attempting to generate check instances that catch the exploitation of nook instances, primarily increasing the specification. One more method is to each generate a system and a proof for its correctness collectively, for which we have now discovered some early success, however extra must be achieved to solidify the method. Additional, it stays to be seen what’s the easiest way to solicit human-written specs for a system—can this be achieved in an iterative, human-in-the-loop method, versus a one-shot, incomplete one. Certainly, human-written specs are incomplete even for manually authored software program, so one would anticipate that future brokers which are extra aligned will more and more train higher judgement when making design selections.
One Doable Knowledge System Synthesis Pipeline [From Here]
Different questions right here contain testing whether or not ranging from a mature system (e.g., Postgres) and eradicating parts/performance can result in increased efficiency or extra consumer belief. Individually, is there a possibility to make the design composable, comprising varied verified parts which are combined and matched given a workload? For instance, maybe the workload hasn’t modified sufficient for the storage layer to be up to date, however maybe the question optimizer requires modifications. A maybe extra viable proposition includes using brokers coupled with proof programs to focus on essential components of the code related to formal proofs, quite than doing so for the whole system.
A remaining alternative right here is to maneuver away from the normal knowledge programs stack with clearly-defined interfaces (e.g., parser, question optimizer, storage supervisor, …) — that had been every largely the prerogative of a single human workforce to handle. As an alternative, brokers can discover new methods to “mix” these parts collectively, maybe figuring out new optimization alternatives consequently. Brokers may also fill in lacking gaps in performance to make present programs rather more feature-complete, or attain feature-parity with different competing programs—or analogously, repeatedly refining open-source programs in response to characteristic requests or points (maybe filed by different brokers!) Doing so in a method that prioritizes correctness, long-term upkeep, and human interpretability might be a problem.
Trying Additional Forward
Within the period of near-free intelligence, knowledge programs matter greater than ever. As brokers tackle the majority of data work, the workload for knowledge programs will change, the substrate they should run on must be constructed, and more and more, they are going to take part in designing knowledge programs themselves. Every of those shifts opens up a brand new, thrilling analysis agenda.
Co-Evolution of Knowledge Methods and Brokers
Trying additional out, the boundaries between brokers and knowledge programs will possible begin to blur. As an example, brokers could design the information programs they themselves run on, defining each the interfaces in addition to the system parts beneath. Each the interfaces and internals could be advanced over time by brokers in a type of recursive self-improvement. There may be additionally a possibility to rethink knowledge programs as a holistic supply of fact for everything of related state: together with uncooked knowledge, reminiscence, and coordination state, additional erasing the distinctions between the information that’s being queried by brokers and knowledge generated on account of agentic exercise. Lastly, knowledge programs could themselves incorporate agentic parts, basically evolving from passive computation engines into clever, proactive, self-optimizing architectures. It’s arduous to foretell what the longer term could maintain. We’re in for a wild experience!
Acknowledgments
The attitude and ongoing work described on this publish are the product of joint analysis and plenty of discussions with fantastic collaborators on the EPIC Data Lab, Data Systems & Foundations group, and the broader Berkeley AI-Methods group. Thanks all!
BibTex for this publish:
@misc{intelligence-is-free-blog,
title={Intelligence is Free, Now What? Knowledge Methods for, of, and by Brokers},
writer={Aditya G. Parameswaran and Shubham Agarwal and Kerem Akillioglu and Shreya Shankar
and Sepanta Zeighami and Rishabh Iyer and Matei Zaharia and Alvin Cheung
and Natacha Crooks and Joseph Gonzalez and Joseph Hellerstein and Ion Stoica},
howpublished={url{https://bair.berkeley.edu/weblog/2026/07/07/intelligence-is-free-now-what/}},
12 months={2026}
}
