A Berkeley research group argues the next decade's load bearing engineering problem is no longer model capability but data systems that coordinate, verify, and remember across thousands of agents.
A GPT-4-class model that cost about $30 per million tokens to run in early 2023 now costs under $1, with some providers pricing below $0.10, according to a new perspective post from UC Berkeley computer scientists led by Aditya G. Parameswaran. Across benchmarks, inference prices have fallen between 9x and 900x per year, with a median decline near 50x per year, per Epoch AI's running dataset.
That collapse is not a forecast. It is the present-tense condition of the field, and it is what makes Parameswaran's argument urgent. If intelligence is cheap enough to be metered like electricity, the load-bearing engineering problem moves down the stack, into three kinds of data systems: those built to serve agents, those agents produce as they work, and those agents generate as their primary output. Parameswaran and his collaborators label these three categories "for, of, and by" agents, and they argue the data layer is the part that has not kept up.
"For" agents means shared memory, retrieval pipelines, tools, and orchestration layers that thousands of reasoning loops can call into without colliding. "Of" agents means the intermediate traces, cleaned intermediate artifacts, and code patches an agent leaves behind. "By" agents means the documents, models, dashboards, and pipelines that ship to humans at the end of a run. Each of the three has different storage, indexing, provenance, and trust requirements.
Berkeley's SKY Lab has begun to formalize this with MAST, a Multi-Agent System Failure Taxonomy, cataloging how pipelines break when coordination slips, when one agent's output becomes another's input, and when verification is bolted on after the fact rather than designed in. Practitioner write-ups on multi-agent architecture patterns, operational-intelligence deployments, and what breaks when multi-agent systems scale describe the same pressure from the operator side. Error compounds across calls, shared state degrades, and the cost of debugging a thousand-agent run is no longer the model's bill but the engineering hours spent tracing which call poisoned which downstream artifact.
That cost is invisible in the per-token price tag. Parameswaran's post is careful to call "free intelligence" a chosen metaphor rather than a settled condition; knowledge work is not all work, and there is no consensus that the for/of/by taxonomy describes what most teams are actually shipping. The argument is that it should, because the alternative is paying commodity inference prices for pipelines that ship unverified output.
A team that swaps a single agent for a coordinated ten-agent pipeline can see its inference bill fall tenfold while its post-hoc audit workload rises by more, because every intermediate handoff becomes a verification surface. The bottleneck has not disappeared; it has moved to the part of the system that remembers what each agent did, that proves a final artifact was assembled from the inputs it claims, and that lets a human reviewer rewind a failed run without replaying it from scratch.
Parameswaran's research directions, including agentic speculation, structured memory for long-horizon pipelines, and synthesizing custom data systems from scratch, all presuppose the model itself is not the bottleneck. Epoch AI's open benchmark-efficiency dataset will keep tracking whether the cost curve continues. The harder measurement, which Berkeley is only starting to formalize, is how often multi-agent systems produce output a human can actually trust without redoing the work.