Anyone who has tried to reason carefully about LLM-based agent architectures will have noticed a persistent problem: the vocabulary is a mess. Industry guides from Anthropic, Google, and LangChain describe systems in terms of data flow and structural topology, while cognitive science literature focuses on what agents actually do cognitively. These two descriptive traditions have developed largely in parallel, and neither is sufficient on its own. The result is that architecturally distinct systems end up sharing the same name, or the same architecture gets described in incompatible ways depending on who is writing about it.
This paper by Huang and Zhou, available at arXiv:2605.13850, proposes a principled solution: a two-dimensional classification matrix that treats cognitive function and execution topology as orthogonal axes, each independently necessary to characterise an agent system.
The Core Argument
The central motivating observation is clean and worth dwelling on. An Orchestrator-Workers topology, one of the most commonly discussed structural patterns, can implement at least three fundamentally different agent behaviours: Plan-and-Execute, Hierarchical Delegation, and Adversarial Verification. These are not minor variations. They have different failure modes, different trust assumptions, and different design trade-offs. Yet under a purely topological classification, they look identical. Conversely, a purely cognitive classification tells you what the agent is trying to do but says nothing about how computation and control actually flow through the system.
The proposed framework resolves this by insisting that both axes are required. The Cognitive Function axis comprises seven categories: Context Engineering, Memory, Reasoning, Action, Reflection, Collaboration, and Governance. The Execution Topology axis comprises six structural archetypes: Chain, Route, Parallel, Orchestrate, Loop, and Hierarchy. The resulting 7x6 matrix yields 42 possible cells, of which the authors identify 27 named patterns, 13 of them original contributions.
Methodology and Validation
The authors demonstrate orthogonality through systematic cross-axis analysis, which is the right thing to do if you want to claim the two axes are genuinely independent rather than just two ways of describing the same thing. Eight representative patterns are defined in detail, giving enough specificity to make the taxonomy actionable rather than merely descriptive.
Validation is conducted across four real-world domains: financial lending, legal due diligence, network operations, and healthcare triage. This is a sensible selection. These domains differ substantially in their operational constraints, including time pressure, the cost of errors, the authority level required for actions, and transaction volume. Using them as test cases allows the authors to ask not just whether the framework can describe existing systems, but whether it predicts anything about architectural choices.
The answer, to the paper's credit, is yes. The cross-domain analysis yields five empirical generalisations the authors call "laws of pattern selection". These govern the relationship between environmental constraints and architectural choices. For example, high failure-cost asymmetry (where false negatives are much worse than false positives, as in healthcare triage) appears to drive adoption of Adversarial Verification patterns rather than single-path reasoning chains. Time pressure, conversely, tends to push systems toward parallel execution topologies even when that introduces coordination overhead.
What This Gets Right
The framework is explicitly model-agnostic and framework-neutral, which matters practically. A classification scheme that only works for LangChain graphs or only applies to GPT-4-based systems would have limited shelf life. The authors have been careful to pitch the vocabulary at a level of abstraction that should remain useful as the underlying models and tooling continue to shift.
The 13 original pattern names are a genuine contribution, not just relabelling. Adversarial Verification, for instance, names something that practitioners implement regularly but that existing taxonomies handle awkwardly, usually folding it into generic "multi-agent" categories that obscure the adversarial structure.
The empirical laws are the most intellectually interesting part of the paper. Moving from a descriptive taxonomy to predictive principles is a significant step, and the five laws provide testable hypotheses. Whether they hold across a broader sample of real deployments is an open question, but framing them as laws rather than observations invites exactly that kind of follow-up work.
Limitations and Open Questions
The paper is ten pages covering a 7x6 matrix, four domains, and five empirical laws. Something has to give, and what gives here is depth on individual patterns. Eight of the 27 patterns are defined in detail; the remaining 19 are named but not fully specified. For a framework intended to provide a "principled vocabulary", this is a gap. Practitioners trying to apply the taxonomy to novel systems will need more than a name and a matrix cell.
The validation methodology, while sensible in domain selection, relies on the authors' own analysis of those domains rather than independent classification by practitioners. Inter-rater reliability is not reported. Given that the framework's utility depends on different people applying it consistently, this is worth addressing in follow-up work.
The Governance category on the cognitive axis is underdeveloped relative to the others. As agent systems take on higher-stakes actions with greater autonomy, governance patterns (covering oversight, auditability, and constraint enforcement) are likely to become central design concerns rather than an afterthought. The framework names this category but does not give it proportionate treatment.
There is also a question about dynamics. The matrix classifies static architectural choices, but real agent systems often shift their effective topology at runtime, for instance moving from a Chain to a Loop pattern when a reasoning step fails and triggers retry logic. How the framework handles architectures that are topologically polymorphic at runtime is not addressed.
Implications
Despite these limitations, this is a useful paper. The field genuinely lacks a shared vocabulary that bridges the industry and academic traditions, and the two-dimensional framing is the right structural insight. The empirical laws, even if preliminary, give researchers something concrete to test and extend. For anyone designing or evaluating LLM-based agent systems, the framework provides a more principled starting point than the ad hoc pattern names currently in circulation.
The most productive next step would be a larger-scale validation study in which practitioners independently classify real deployed systems using the taxonomy, with inter-rater agreement as a primary metric. That would tell us whether the framework is not just logically coherent but practically usable.
Full paper: arXiv:2605.13850