Agentic state-of-the-art search: the survey that argues with itself

A state-of-the-art search should tell you what a field has already solved before you spend money solving it again. Agentic state-of-the-art search changes how that answer is built: the system queries, reads, revises, and queries again. The difference is not stylistic. It shows up in recall, and recall is what separates a good survey from a confident one.

The corpus outgrew the query

WIPO's World Intellectual Property Indicators 2025 counts a record 3.7 million patent applications in 2024, up 4.9% on 2023 and the fifth consecutive year of growth. Filings have risen 1.9-fold since 2010. Asia's share climbed from 60% in 2014 to 70.1% in 2024, with China alone accounting for 49.1%.

The non-patent side is no smaller. The US National Science Board puts worldwide science and engineering publication output at 3.3 million articles in 2023, using Scopus. In most technology areas the relevant prior work is split between the two streams and written in several languages.

View data

Origin	Applications
China	1.8e+06
United States	501831
Japan	419132

Even one office is multilingual. The EPO received 201,974 applications in 2025; 57% originated outside its member states, with China (10.9%) overtaking Japan (10.5%) as the third-largest source and Korea at 7.1%.

View data

Label	Value
EPO member states	42.9%
United States	23.3%
China	10.9%
Japan	10.5%
Korea	7.1%
Others	5.3%

A keyword-and-classification search built by one person in one afternoon samples that field along a single axis and calls the sample a landscape.

What "agentic" changes

An agentic search issues a query, reads what comes back, decides what it has learned, and issues a different query. The loop repeats across angles until the evidence stops improving. Two recent studies of long-horizon search agents show why that is not a small change in mechanics.

Diagnosing Search Behavior and Failure Modes in Long-Horizon Search Agents (arXiv 2608.01913) compared six agents on the BrowseComp-Plus benchmark with the retriever held fixed. Answer accuracy was only weakly related to search effort. What it tracked was cumulative retrieval recall: the share of relevant evidence the agent's queries eventually surfaced. On the same question, the agent that retrieved the gold evidence answered correctly most of the time; the agent that missed it almost never did. The best agents also issued far fewer redundant queries.

Bridge Evidence: Static Retrieval Utility Does Not Predict Causal Utility in Multi-Step Agentic Search (arXiv 2607.15253) explains where the extra recall comes from. Over 1,000 multi-hop questions, the correlation between a document's static relevance score and its causal usefulness to the agent was −0.026, effectively zero. About a third of the documents the agent read looked worthless to a stateless ranker yet were load-bearing: they carried the entity or term that made the next query possible. Terms that discriminated relevant from irrelevant candidates reached the agent's next query 4.02 times more often than terms found only in noise.

For a survey this is the whole game. The expired 2009 patent that solved your problem under a different name is invisible to a query built from your vocabulary. It becomes visible once an agent has read a review article using both vocabularies and pivots on the shared term.

Why the landscape report underdelivers

The traditional answer is a landscape report from an analytics vendor. Published pricing puts a scoped project at roughly $5,000 for a narrow single-technology review and $50,000 or more for a broad multi-jurisdiction landscape; specialist assignments are quoted from $10,000 to well over $100,000.

Cost is not the main issue. The report is commissioned by a strategy team, built on classification codes and assignee counts, and delivered as slides. It rarely reaches the people drafting and filing, and it was never structured to answer the questions the business asked. ipCapital Group notes that patent filings typically lead product announcements by 18 to 36 months — exactly the signal a landscape should surface, but only if the documents were read rather than binned by code.

What an agentic survey looks like in practice

Esgenix runs a state-of-the-art search as a workflow the user directs, not a report the user receives.

1. 1

   Describe the problem space

   The technology or research question, in plain language.

2. 2

   Set the business questions

   Three to five questions the survey must answer; the report cannot finalise while any remains open.

3. 3

   Survey in rounds

   Agents search 220 million patents across 108 jurisdictions and more than 100 million research papers semantically, casting wide enough to include expired and foreign art, in batches a reviewer can read.

4. 4

   Approve the taxonomy

   The technology map is renamed, merged and approved by the user before any conclusion is drawn on it.

5. 5

   Build the analysis

   A problem ledger of what is solved and what remains open, player signals with build, buy, partner, avoid or watch calls, technology evolution and assignee concentration.

6. 6

   Close the report

   Business questions answered in writing, reviewed and approved, with every chart linked to its underlying references.

Three design choices do most of the work.

The business-question gate. The survey is framed by the questions the team needs answered, and the report will not finalise while one is open. The deliverable becomes "what you asked" rather than "data about the field".

The taxonomy gate. Agents propose a technology taxonomy from what they found; a human renames, merges and approves it before analysis begins.

Linked evidence. Click any chart segment and the underlying references filter everywhere. One click between a headline and its evidence is what makes the report defensible in front of a board or an examiner.

The honest version of the "80% rule"

Patent-information marketing has repeated for decades that "80% of technical information is found only in patents". A 2016 review in World Patent Information traced the figure to a 1977 USPTO Technology Assessment and Forecast report built on a small sample, and concluded the phrase cannot be verified as a general claim. The careful evidence is field-specific: a 2005 study of chemistry literature found that 95% of substances disclosed in patents did not appear in the non-patent literature.

That is the concrete argument for agentic search over a well-built Boolean query. Agents are not cleverer readers. They keep reading after the first hit, pivot on what they learn, and are measured on recall rather than on how confident the first page looked.

What to do before you commission a survey

• Write the three to five business questions first, and refuse a deliverable that leaves any of them open.
• Ask how recall is measured, not how many searches or documents the run produced.
• Require both streams: patents and research papers, in every language, including expired and foreign art.
• Approve the taxonomy yourself before anyone draws a chart on it.
• Check that every chart segment resolves to the documents behind it in one click.
• Route the thin areas the survey exposes straight into white-space analysis and the next filing session.

The field is now 3.7 million filings a year and growing. The only searches that keep up are the ones that argue with themselves until the evidence stops moving.

Frequently asked questions

What is the difference between a state-of-the-art search and a prior-art search?

A prior-art search asks whether one specific claim is new. A state-of-the-art search maps a whole technology area: what has been solved, by whom, how approaches have shifted, and where the open problems are. It informs filing programmes and R&D direction rather than a single application.

How does agentic search improve recall over a Boolean search?

An agent queries, reads the results, and reformulates on what it learned, so it can pivot on terminology it did not know at the start. Studies of long-horizon search agents show accuracy tracks cumulative recall, and that many decisive documents look irrelevant to a static ranker but carry the terms that unlock the next query.

Why does the number of searches an agent runs not predict a better result?

Because effort and coverage are different things. On BrowseComp-Plus, agents that retrieved the gold evidence answered correctly most of the time and issued fewer redundant queries; agents that missed it almost never recovered by searching harder. Recall is the variable that moves accuracy.

Does an agentic survey replace the analyst?

No. The two gates in the Esgenix workflow — the business questions and the taxonomy — are human decisions, and the final report is edited and approved by the team. Agents do the reading at scale; people decide what the map should look like and what the answers mean.

What corpus does the Esgenix survey cover?

Patent collections spanning 108 jurisdictions, around 220 million documents, plus more than 100 million research papers, searched semantically. Expired and foreign art is included deliberately, because for state of the art the expired patent that solved the problem is often the whole answer.

How current is the result?

The survey runs against the live corpus at the time it is ordered. If a player or a sub-area needs a deeper pass, an additional round runs before the report closes rather than waiting for a next-year refresh.