The 28-Pattern Matrix
A 7×6 framework for agent architecture.
The model spends. The harness budgets. This catalog is the vocabulary you can put in your project tomorrow.
Most "agent architecture" guides give you a flat list — Reflection, ReAct, Multi-Agent, Tree of Thoughts, Reflexive Metacognitive, and so on. A flat list answers what patterns exist. It does not answer where my problem sits, and which pattern lives at that coordinate.
A loan-evaluation agent crashes not because Reflection is missing, but because Perception-stage budget allocation dropped the disqualifying document. A multi-agent code reviewer drifts not because ReAct is wrong, but because Collaboration topology without conflict resolution allows contradictory edits to accumulate. Knowing which axis the failure sits on matters more than knowing the failing pattern's name.
02-03) reference the chapter introducing the pattern in Designing AI Agents.
Two axes
Cognitive function (the rows) — what kind of work the agent is doing. Seven categories: Perception, Memory, Reasoning, Action, Reflection, Collaboration, Governance.
Execution topology (the columns) — what shape the work takes. Six archetypes: Chain, Route, Parallel, Orchestrate, Loop, Hierarchy.
Crossed, the axes form a 7×6 = 42-cell matrix. Twenty-eight cells are populated by named patterns, some borrowed from existing literature (RAG, Chain-of-Thought, Plan-and-Execute), some named here for cells that had no established name (Context Triage, Failure Journal, Observability Harness, others). Fourteen cells are intentionally empty — representing the industrial blanks and the future-research frontier.
Pattern white papers
Every named cell in the matrix has a one-page enterprise white paper — what problem it solves, what coordinate it sits at, where it tends to break, and what to monitor once it ships. They are written for the whole organization, not just engineers: product, architecture, R&D, compliance, and management.
Following the ADPS Chinese-first publishing pipeline, the 33 deep-dive white papers are published in Chinese first; English editions are in progress. Browse the 33 white papers (Chinese) →
The matrix
The full matrix, runnable code, and an engineering slice for each pattern live in the companion GitHub repository:
- huangjia2019/agent-design-patterns — standalone catalog, each pattern self-contained and runnable in isolation. (Will migrate under the ADPS organization.)
- huangjia2019/designing-ai-agents — companion to the Manning book; Argus runtime grows chapter by chapter.
The methodology
The matrix becomes operational through the Pattern Selection Card — a three-panel decision tool:
- ASSESS — what cognitive functions does the task need?
- ROUTE — what topology suits the constraints?
- SELECT — which named patterns sit at those intersections?
The card turns "I'm building an agent for X" into a candidate pattern list in under fifteen minutes.
Where to read more
- The position paper: arXiv:2605.13850
- The Manning book: Designing AI Agents
- The runnable catalog: agent-design-patterns