… authorities of the folks, by the folks, for the folks …
— Abraham Lincoln, Gettysburg Tackle (1863)
The price of AI is dropping quickly. GPT-4-class capabilities price roughly $30 per million tokens in early 2023; immediately the identical runs beneath $1, and some suppliers are pushing prices under $0.10. Throughout benchmarks, inference costs have fallen between 9x and 900x per yr, with a median decline close to 50x. Even frontier fashions are getting dramatically cheaper every era, with open-source fashions following intently 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 information work is right here immediately, and getting cheaper by the month. At this charge, we’re quickly coming into the period of just about free intelligence—the type that’s greater than sufficient for on a regular basis data 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 together with his collaborators. It’s half panorama survey and half perspective, and several 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 imagine three new challenges—and alternatives—stem from near-zero inference prices:
Knowledge Methods For Brokers. Brokers will quickly grow to be 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 information work, a brand new substrate is required for hundreds 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 seem like?
Knowledge Methods By Brokers. Brokers are quickly turning into able to synthesizing complete knowledge programs in a single go—that means we will rebuild customized programs for every new workload. Verifying that such programs match meant habits 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 knowledge 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 software. It performs what we name agentic hypothesis: 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 area, every person request may quantity to 1000s of particular person SQL queries. Now, customers can difficulty ‘high-level’ knowledge duties, e.g., root-cause evaluation—e.g., ‘why did espresso gross sales in Berkeley drop this yr’—or exploratory cohort evaluation—e.g., ‘which person segments are almost certainly to churn subsequent quarter’—every involving a combinatorial area of potential joins, aggregations, and filter combos.
Knowledge Methods Redesigned to Extra Successfully Assist Agentic Hypothesis
The requests from these brokers have numerous alternatives for optimization. As an example, on a text-to-SQL benchmark with a number of brokers trying every process, solely 10-20% of the sub-plans are distinct. Thus, 80-90% of sub-queries carry out duplicate work. The identical experiments present process success charges considerably growing 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 quicker. It may 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 remaining is critical or useful.
One other alternative right here is to rethink the question interface fully: as an alternative of brokers issuing a single SQL question at a time, they may as an alternative difficulty a batch of queries, every with its personal approximation necessities. Since enumerating an exponential search area (as within the root trigger or cohort evaluation examples above) isn’t a superb use of agentic reasoning potential, maybe knowledge programs ought to assist higher-level primitives moderately than requiring brokers to record every SQL question explicitly. One thought right here is to attract on DBT-style Jinja macros to supply looping-based primitives for brokers to work together with knowledge programs.
A Caffeinated Military of Brokers Able to Tirelessly Full Your Knowledge Duties
A ultimate alternative right here is to cease considering of knowledge programs as passive executors of queries; knowledge programs may very well be proactive, as they possess extra grounding in knowledge and system traits that brokers might lack a priori—they may steer brokers in several instructions, present outcomes for associated queries, and likewise present performance-level suggestions (e.g., as an alternative of executing an costly question, the system may first present the agent a latency estimate). The explanation 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 end result. In reality, the information system may additionally put together each materialized and digital views for an agent prematurely, offered to the agent as a part of context, as this can be cheaper or more practical than having an agent creator or use them.
Knowledge Methods Of Brokers
Beforehand, we targeted on how brokers work together with knowledge programs. Now, we contemplate every thing else brokers must preserve working: the place they reside, 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. Nevertheless, the inference stack itself is being abstracted away via APIs (e.g., from OpenAI or Anthropic), or, for open-weight fashions, via serving frameworks that disguise low-level particulars. To this point, the agentic substrate has been managed via harnesses like Claude Code and Codex, coupled with numerous mechanisms to retailer and retrieve reminiscence.
First, on the reminiscence entrance, the present knowledge is that recordsdata are all you want; brokers write to unstructured markdown (MD) recordsdata, which may then be searched utilizing grep, or by way of embedding-based retrieval. In reality, many argue that the answer to continuous studying is having brokers devour so much (e.g., a whole codebase, slack, firm wikis, …) after which write their learnings into MD recordsdata, that are then retrieved selectively on demand. Certainly, file programs, bash scripting, and MD recordsdata are and can nonetheless be essential for brokers. Nevertheless, at scale, when brokers are doing the overwhelming majority of information work, this strategy will now not be efficient.
Given restricted context home windows, retrieving all MD file fragments which may be related and stuffing it into the context will break down in some unspecified time in the future. Even when context home windows proceed to develop, there are latency advantages to not put all data into context — and in lots of circumstances, e.g., when data work includes interacting with massive databases or code bases, it will likely be infeasible to serialize all related knowledge into context.
Knowledge Methods As A Substrate for Multi-Agent Swarms
One may use a data graph illustration, however data graphs endure from the identical limitations as unstructured MD-based reminiscence attributable to 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 aspects) of curiosity. For instance, an agent debugging a flaky take a look at ought to have the ability to pull solely the reminiscences tagged with the related module, language, framework, and failure mode—moderately retrieving primarily based on key phrases or embedding similarity. A separate difficulty is what to really retrieve; uncooked agent traces with errors aren’t very helpful as they’ll induce brokers to repeat the identical mistake—as an alternative, we wish the retrieved reminiscence to be corrective.
We just lately explored a associated notion of structured reminiscence, the place we arrange reminiscence throughout numerous attributes, every of which may very well be set as * to point common applicability, or set as a listing of values to be matched. For an information agent, the scale may embrace the columns and tables, kind of operation, and eventually, open-ended natural-language corrective directions. So, we may embrace reminiscence that solely applies to a given kind of operation (e.g., ‘when performing date-time operations, use fiscal yr versus calendar yr conventions’), or a given desk (e.g., ‘column product_cleaned is most well-liked over column product when querying on product title’). One open query is defining an application-specific structured reminiscence—or what others have referred to as world fashions for reminiscence. We imagine that is akin to defining a schema for every utility—and maybe brokers themselves might help us outline and refine it over time.
One Attainable Method To Retailer and Retrieve Structured Information [From Here]
Structured reminiscence will likely be helpful additionally for evolutionary frameworks to successfully handle search areas. Certainly, storing, structuring, and mining massive volumes of single and multi-agent traces might help future brokers grow to be way more environment friendly—probably enabling efficient recursive self-improvement via structured memory-based mechanisms.
One other problem is to assist concurrent edits to shared reminiscence, and concurrent edits usually, when there are numerous 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 methods will suffice when hundreds of brokers are trying to edit shared state on the identical time. As an example, when brokers try numerous potential transactions in response to a person request, the results of the overwhelming majority of those transactions have to be rolled again—with solely the one ‘appropriate’ transaction’s end result persisting. Work on supporting exactly-once semantics is related right here, as are underlying methods primarily based on CRDTs and operational transformation. For updates to fuzzy mechanisms similar to reminiscence, we could possibly sacrifice on consistency for good correctness within the curiosity of latency. Whereas brokers can purpose about semantics to compensate or roll again their actions to finally finalize most duties, the first problem lies within the diploma to which they step on one another’s toes throughout the course of. An essential failure mode to be prevented is a type of “livelock,” the place incessant compensating actions stop any significant progress.
Past shared state, different considerations emerge when making an attempt to assist a military of brokers, together with what to do when brokers fail, how brokers ought to talk with one another (straight or via 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, similar to Temporal, however it stays to be seen if such options will apply at scale throughout hundreds 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 may contain iterative dialogue and compromise; for agentic swarms, we should outline the mechanisms that enable 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 will likely be essential. It stays to be seen if that is finest carried out by way of centralized coordination, or if a decentralized strategy is critical.
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 an entire, workload-specific analytical engine—in minutes to some 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 shops from scratch, focused to the workload. In reality, fashionable IDEs, similar to Kiro, elevate specs for programs improvement to be a first-class citizen.
Brokers Can Synthesize Customized Knowledge Methods From Scratch
The principle difficulty, nonetheless, is that specs are sometimes imperfect, and don’t cowl all nook circumstances. Current-day brokers will exploit the lacking specs to reward-hack their method to a excessive efficiency metric. In our customized key-value retailer work, we discovered that one method to alleviate that is to have auxiliary verification brokers making an attempt to generate take a look at circumstances that catch the exploitation of nook circumstances, basically increasing the specification. Yet one more strategy 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 carried out to solidify the strategy. Additional, it stays to be seen what’s one of the best ways to solicit human-written specs for a system—can this be carried out 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 choices.
One Attainable 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 larger efficiency or extra person belief. Individually, is there a possibility to make the design composable, comprising numerous verified parts which are blended 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 crucial elements of the code related to formal proofs, moderately than doing so for the complete system.
A ultimate 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 have been every largely the prerogative of a single human group to handle. As an alternative, brokers can discover new methods to “mix” these parts collectively, maybe figuring out new optimization alternatives consequently. Brokers can even fill in lacking gaps in performance to make current programs way 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 will likely 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 information work, the workload for knowledge programs will change, the substrate they should run on must be constructed, and more and more, they’ll 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 seemingly begin to blur. As an example, brokers might design the information programs they themselves run on, defining each the interfaces in addition to the system parts beneath. Each the interfaces and internals might 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 reality 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 might themselves incorporate agentic parts, essentially evolving from passive computation engines into clever, proactive, self-optimizing architectures. It’s arduous to foretell what the longer term might maintain. We’re in for a wild trip!
Acknowledgments
The angle and ongoing work described on this publish are the product of joint analysis and lots of discussions with fantastic collaborators on the EPIC Knowledge Lab, Knowledge Methods & Foundations group, and the broader Berkeley AI-Methods neighborhood. Thanks all!
BibTex for this publish:
@misc{intelligence-is-free-blog,
title={Intelligence is Free, Now What? Knowledge Methods for, of, and by Brokers},
creator={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/}},
yr={2026}
}
