When Research Becomes a Tree: Why Static-DRA Matters in an Agentic World

A research agent enters a company budget meeting.

That sounds like the beginning of a bad consulting joke, but it is exactly where “deep research” systems are heading. The first generation of excitement was about capability: can an AI agent search, plan, decompose, synthesize, and write a report that feels less like a chatbot answer and more like an analyst memo? Fine. The next question is less glamorous and far more operational: can the company control how much research the agent performs before the invoice becomes a small weather event?

This is where Static-DRA is interesting. The paper, A Hierarchical Tree-based Approach for Creating Configurable and Static Deep Research Agent (Static-DRA), proposes a deep research agent whose central selling point is not maximum autonomy but configurable restraint.¹ It gives users two explicit controls: Depth, meaning how many levels the agent may recursively decompose a topic, and Breadth, meaning how many subtopics it may generate at each branching point.

That may sound almost too simple. In the current agentic AI mood, “static” can feel like an insult, the architectural equivalent of calling a product “manual” in a room full of automation vendors. But the paper’s useful idea is precisely that a static workflow can be valuable because it is static. It turns open-ended research into a bounded tree. Less dramatic than a self-replanning autonomous agent, yes. Also less likely to wander into a token-burning philosophical expedition because someone asked for “a quick overview.”

The business lesson is not that Static-DRA defeats frontier commercial deep research products. It does not. The useful lesson is that research automation needs a visible control surface. Static-DRA offers one.

Static-DRA is a research machine with a brake pedal

The paper positions Static-DRA against two backgrounds.

The first is traditional retrieval-augmented generation. Classic RAG retrieves a set of documents and asks a model to generate an answer from them. That works for narrow questions, but it is a brittle fit for research tasks that require decomposition, comparison, multi-hop exploration, and synthesis. A single retrieval pass often gives the model a pile of context, not a research plan.

The second background is dynamic deep research agents. These systems can plan, search iteratively, revise their queries, and decide what to investigate next. They are powerful, but they also create governance problems. A dynamic agent may be harder to budget, harder to audit, and harder to explain because its search path emerges during execution.

Static-DRA takes a different bargain. It does not promise full adaptive freedom. It says: give me a maximum depth, give me a maximum breadth, and I will run a hierarchical research process inside those limits.

That bargain matters because enterprise AI buyers rarely purchase “autonomy” in the abstract. They purchase workflows that can be priced, monitored, reviewed, and defended when something goes wrong. A research tree is not as fashionable as an autonomous planner. It is, however, much easier to draw on a whiteboard without everyone pretending the squiggle is a strategy.

The mechanism: Supervisor, Independent Agent, Worker

Static-DRA is built around three agent roles. The names are plain, which is a mercy.

The flow is simple. A research topic arrives. The Supervisor checks whether the current depth still allows decomposition and whether the topic can be meaningfully split. If yes, the Independent Agent generates subtopics. Each subtopic gets its own Supervisor. If no, the Worker researches the topic directly.

This structure gives Static-DRA its main governance feature: research is not an unbounded conversation; it is a bounded expansion process.

The paper also adds two practical controls that deserve attention.

First, the system keeps track of past research topics and runs a duplicate-topic sanity check before executing a worker. This matters because recursive decomposition can easily produce overlapping subquestions. If two branches ask nearly the same thing, the agent burns tokens while pretending to be thorough. The duplicate check is a small but useful guardrail.

Second, the system stores past citations and generated reports in a shared state. The final report is assembled from worker outputs while preserving a table of contents, report body, and citation section. In other words, the research tree also becomes the report outline. That is a useful design pattern: the execution trace doubles as the explanation artifact.

Depth and breadth turn research effort into a budget language

The paper’s core contribution is the pair of user-facing parameters.

Depth controls how many decomposition levels the system may explore. A depth of 1 produces shallow decomposition. A depth of 2 allows subtopics to become deeper subtopics. Higher depth means more opportunities for multi-hop exploration.

Breadth controls the number of subtopics the system may generate at a branching point. A breadth of 5 tells the system to look for up to five independent subtopics, assuming the topic can actually be split that way.

The important phrase is “up to.” The paper makes clear that breadth is constrained by the topic itself. If a topic naturally decomposes into only three independent parts, setting breadth to five does not magically create five useful branches. It creates, at most, the number of meaningful branches available. This is a quiet but important correction to the common “more branches equals more intelligence” fantasy. Sometimes more branches just means more synonyms wearing a fake mustache.

The paper also reduces breadth at deeper levels. The prose describes this as reducing breadth by a factor at each deeper level, while the algorithm and appendix formula operationalize it as subtracting 2 from breadth at each level. The appendix gives the maximum number of worker topics as:

Here, $d$ is configured depth, $b$ is configured breadth, and $ns(d)$ is the maximum number of lowest-level research topics sent to workers. The max(..., 1) term prevents the branch count from collapsing below one as depth increases.

For business readers, the exact formula is less important than the operating principle: depth and breadth define an upper bound on LLM work. A configuration with $d=2$ and $b=5$ has a larger possible worker set than $d=1$ and $b=2$. The system may not hit the maximum if the topic cannot be split enough, but the user has a visible ceiling.

This is the paper’s strongest practical idea. Many agent workflows sell intelligence as an open-ended loop. Static-DRA sells it as a tree with known knobs. That is less romantic. Finance departments are not paid to be romantic.

The parameter tests show scaling, not magic

The paper includes a configuration comparison using the topic: “What are the investment philosophies of Duan Yongping, Warren Buffett, and Charlie Munger?” It tests three configurations:

The likely purpose of this test is sensitivity analysis: it shows how the depth-breadth configuration changes report size, number of subtopics, and quality score. It is not the main leaderboard evidence, and it should not be read as a universal scaling law.

The pattern is still useful. More depth and breadth generated more subtopics, a much larger report, and a higher overall score. But the relationship is not a clean “pay twice, get twice the quality” curve. Moving from d2_b3 to d2_b5 expands the report from 11.3 KB to 74.5 KB, while the overall score rises from 0.35 to 0.41 in the figure. That is a meaningful increase, but not a free lunch. It is more like paying for the whole buffet and discovering the marginal utility of the fifth plate.

This is exactly the kind of evidence enterprises should want before deploying research agents. Not because the numbers transfer directly to their workflows, but because the system makes the trade-off measurable. Research intensity becomes a tunable choice rather than a vague instruction like “be comprehensive.”

The benchmark result is respectable, not frontier-dominating

For the main evaluation, the paper uses DeepResearch Bench and reports RACE scores. RACE, or Reference-based Adaptive Criteria-driven Evaluation, evaluates the quality of final generated research reports across criteria such as comprehensiveness, insight, instruction following, and readability.

Static-DRA is evaluated with gemini-2.5-pro, depth 2, and breadth 5. It achieves an overall RACE score of 34.72.

The comparison table is important because it puts the result in context:

The honest interpretation is straightforward. Static-DRA does not catch specialized deep research products. It sits well below Gemini Deep Research, OpenAI Deep Research, and Claude Research in the reported RACE leaderboard comparison. But it performs above several more general search or model baselines in overall score.

That is a reasonable result for the paper’s thesis. Static-DRA is not presented as the final boss of research agents. It is a configurable architecture that can make a strong base model behave more like a structured research system. The improvement is especially visible in comprehensiveness and insight relative to some general baselines, though its insight score remains far below frontier deep research systems.

The result also helps avoid a lazy misconception: static does not mean useless. A static tree can still create useful multi-hop research behavior. It just does so through preset structure rather than continuous adaptive replanning.

Read the evidence by purpose, not by excitement level

The paper contains several kinds of evidence, and mixing them together would make the article sound more certain than the paper actually is. A clean reading looks like this:

This matters because agent papers often invite overreading. A leaderboard table becomes a product claim. A parameter test becomes a law. A topic heatmap becomes a market segmentation strategy. Please do not do that. The paper is useful enough without giving it a cape.

The missing FACT evaluation is not a footnote; it changes deployment meaning

DeepResearch Bench includes two evaluation frameworks: RACE for final report quality and FACT for factual abundance and citation trustworthiness. Static-DRA is evaluated using RACE. The paper does not provide FACT results.

That boundary matters for enterprise use.

A research report can be readable, comprehensive, and well-structured while still having weak citation discipline. Static-DRA’s Worker nodes use Tavily web search, filter results using a relevance threshold, and preserve cited URLs. This is a reasonable retrieval-and-citation pipeline. But preserving URLs is not the same as proving that claims are faithfully supported by sources.

For business deployment, this distinction is critical.

If Static-DRA is used for internal landscape scans, market maps, sales research, or early-stage strategy memos, RACE-style quality may be enough to justify experimentation. If it is used for legal research, investment diligence, medical content, regulatory monitoring, or public-facing reports, citation trustworthiness becomes non-negotiable. The paper’s architecture gives a place to attach citation governance, but the evaluation does not prove that governance.

That is not a fatal flaw. It is a boundary. Static-DRA is a promising scaffold for controlled research automation. It is not, based on this paper alone, an audited fact engine.

The business value is controlled research intensity

Static-DRA’s business relevance is not “replace analysts.” That slogan needs to be retired, preferably in a quiet room with no Wi-Fi.

The better interpretation is: Static-DRA converts research intensity into an explicit operating choice.

A company could imagine three tiers of research workflow:

The value is not only cost control. It is managerial control. A team can decide that a low-stakes task deserves a shallow tree, while a high-stakes task deserves a deeper tree plus human review. The same architecture can support different effort levels without rewriting the workflow from scratch.

There is also an audit benefit. Because the research process is tree-shaped, reviewers can inspect which subtopics were explored, which branches produced worker reports, and which citations were collected. In a dynamic autonomous workflow, the execution path may be harder to reconstruct. In Static-DRA, the tree itself becomes part of the governance artifact.

That may sound boring. Good. Boring is underrated in systems that spend money while speaking confidently.

Where Static-DRA fits best

Static-DRA is most attractive when three conditions hold.

First, the research task can be decomposed into relatively independent subtopics. Comparative research, landscape analysis, market mapping, literature scanning, and “explain this field” reports fit this pattern well. A tree works when the question has branches.

Second, users care about budget and repeatability. If every research run needs a predictable ceiling, static configuration is valuable. This is especially relevant for teams building internal tools on top of commercial LLM APIs where input tokens, output tokens, and repeated calls all affect cost.

Third, the organization can tolerate a structured but not fully adaptive process. Static-DRA may miss opportunities that a more dynamic planner would pursue after discovering unexpected evidence. That is the price of control. Sometimes the price is worth paying.

The system is less naturally suited to tasks where the research direction must change based on intermediate discoveries. Investigative journalism, open-ended due diligence, adversarial intelligence, and complex scientific synthesis may benefit from dynamic replanning. Static-DRA can still help as a scaffold, but its static nature becomes a constraint rather than a feature.

What remains uncertain before production use

The paper gives enough detail to understand the architecture and enough benchmark evidence to justify interest. It does not answer every question a business team would ask before deployment.

The first missing piece is cost measurement. The paper argues that depth and breadth control LLM interactions, and the appendix provides a worker-topic formula. But it does not provide real API cost curves, token consumption distributions, or latency measurements under different configurations. A production team would need those numbers.

The second missing piece is citation reliability. As discussed above, the paper evaluates RACE, not FACT. For many business workflows, especially those involving compliance or external publication, citation trustworthiness is not optional garnish.

The third missing piece is ablation. The paper does not isolate how much value comes from each component: the Supervisor, Independent Agent, duplicate-topic check, breadth reduction, Tavily filtering, or final report assembly. Without ablation, we know the whole system works to a degree; we do not know which parts carry the weight.

The fourth missing piece is human review workflow. Static-DRA produces markdown reports with citations, but the paper does not design an enterprise review layer: approvals, red-flag detection, source grading, claim-level citation mapping, or reviewer feedback loops.

These are not complaints for sport. They are the difference between a research prototype and a deployable research operation.

The real lesson: static can be strategic

The agent world loves dynamism. Dynamic planning sounds intelligent. Adaptive tool use sounds modern. Self-replanning sounds like the future has put on a blazer.

Static-DRA is useful because it challenges that instinct. It says a research agent does not always need to be maximally autonomous. Sometimes it needs to be legible, configurable, and bounded. The paper’s results support that narrower but practical claim: a static hierarchical tree, powered by a strong LLM and web search, can produce meaningful deep research behavior, improve with larger depth-breadth settings, and perform respectably on RACE compared with several general baselines.

The architecture is not a replacement for frontier deep research products. It is more like a design pattern for teams that want to build their own controlled research systems: define the tree, expose the knobs, track the branches, collect the citations, and make the cost-quality trade-off visible.

That is not the flashiest version of agentic AI. It may be one of the more useful ones.

Because in business, the best research agent is not the one that can think forever.

It is the one that knows when the budget says stop.

Cognaptus: Automate the Present, Incubate the Future.