Editorial illustration for New Two‑Dimensional Framework Maps AI Agent Design Patterns
New Two‑Dimensional Framework Maps AI Agent Design Patterns
Agent architectures are proliferating faster than we can name them. Industry guides map data flows; cognitive science surveys map mental functions. Each lens reveals something real, but neither one sees enough.
The same execution topology, say, Orchestrator-Workers, can run three completely different patterns: Plan-and-Execute, Hierarchical Delegation, or Adversarial Verification. Three systems, three failure modes, three design trade-offs, all hidden behind a single flowchart. That ambiguity is not just academic; it’s dangerous when you’re trying to debug, scale, or govern an agent in production.
We need a map that cuts both ways. This paper delivers one: a two-dimensional framework that marries cognitive function, seven categories from Context Engineering to Governance, with execution topology, six structural archetypes from Chain to Hierarchy. The resulting 7x6 matrix names 27 distinct patterns, 13 of them original.
It doesn’t just describe what agents do or how they route data. It tells you why they fail and where to intervene.
Cross-domain analysis yields five empirical laws of pattern selection governing the relationship between environmental constraints (time pressure, action authority, failure cost asymmetry, volume) and architectural choices.
This framework does not pretend to be the final word. It is, instead, a map where the territory is still being drawn. By placing cognitive function and execution topology on orthogonal axes, we surface what the industry guides and the cognitive surveys each miss: that architecture is not just how data flows or what an agent intends, but the interplay between the two.
The Orchestrator-Workers pattern is not one thing. It is three. That distinction matters when failure modes diverge, when a failure in Plan-and-Execute looks nothing like a failure in Adversarial Verification.
The 27 named patterns are not a checklist. They are a grammar. Use it to describe what you build, to see what you have not considered, and to argue about trade-offs with precision instead of vibes.
The field will outgrow this taxonomy; that is its purpose. A good map shows you where to go next.
Related Reading
Policy & Regulation Grammarly faces class-action suit over AI ‘Expert Review’ feature
Policy & Regulation Claude Mythos highlights EU AI safety gaps, says researcher Caroli
Policy & Regulation After ditching AI fitness apps and a Fitbit, I return to Peloton classes
From Archives Achiam said he interrupted Musk's 2018 OpenAI speech to flag AGI safety concerns
From Archives Longer Reasoning Paths Increase Per-Question Position Bias in QA Models
Common Questions Answered
What is the two-dimensional framework for mapping AI agent design patterns?
The framework places cognitive function and execution topology on orthogonal axes to reveal the interplay between how an agent intends to act and how data actually flows through the system. This approach surfaces distinctions that traditional industry guides and cognitive science surveys each miss individually, providing a more comprehensive view of agent architecture design.
Why can the same Orchestrator-Workers execution topology represent three different patterns?
The Orchestrator-Workers topology can run Plan-and-Execute, Hierarchical Delegation, or Adversarial Verification patterns, each with completely different failure modes and design trade-offs despite sharing the same flowchart. This ambiguity demonstrates why execution topology alone is insufficient for understanding agent architecture, and why cognitive function must be considered as an independent dimension.
How does this framework differ from industry guides and cognitive science surveys?
Industry guides typically map data flows while cognitive science surveys map mental functions, but each lens only reveals part of the picture. The two-dimensional framework combines both perspectives by treating execution topology and cognitive function as separate but interrelated dimensions, exposing architectural distinctions that neither approach alone can identify.
Why does the distinction between different Orchestrator-Workers patterns matter for AI agent design?
The distinction matters because each pattern variant has divergent failure modes and unique design trade-offs that directly impact system reliability and performance. Understanding which specific pattern an agent implements, rather than just its execution topology, is critical for anticipating failure scenarios and making informed architectural decisions.
