huangjia2019 / agent-design-patterns

What is agent-design-patterns?

This is a catalog of 28 AI agent design patterns organized into a two-axis matrix—seven cognitive functions (perceive, reason, act, and so on) crossed against six execution topologies (chain, parallel, loop, and so on). The framework gives each pattern a coordinate in the matrix, which answers the practical question of "where does my problem sit and which pattern lives at that coordinate?" Written in Python with runnable code for each pattern, it includes example scenarios and test suites. The patterns draw from production codebases including Claude Code, Aider, OpenHands, and DeerFlow, with real file references cited in each pattern's documentation. It serves as the companion code repository for the Manning book "Designing AI Agents."

Why is it gaining traction?

Most agent architecture guides give you a flat list of patterns without explaining why a pattern belongs in one spot versus another. This matrix approach forces an answer to that question—a pattern at "reason × loop" behaves differently from one at "act × route," and the coordinates make that distinction visible. Each pattern implementation stays small (50-250 lines) with clear invariants, so you can read one in an afternoon and drop it into a project. The production code citations give each pattern credibility beyond toy examples.

Who should use this?

Backend engineers building multi-step agent workflows who need a vocabulary for discussing architecture decisions. Platform teams standardizing agent patterns across products will find the matrix useful for pattern selection. Anyone frustrated with agent tutorials that describe what patterns exist without explaining where they apply. Not for teams looking for a production runtime framework—this is design vocabulary you apply on top of LangGraph, DeerFlow, or OpenHands.

Verdict

The matrix organization is genuinely useful for thinking through agent architecture, and the runnable code is solid for the patterns that are complete. However, with 21 stars and a 1.0% credibility score, this is an early-stage project from an author with a book to sell. Fourteen of 28 patterns are scaffolded placeholders, and collaboration and governance patterns lack code entirely. Worth bookmarking and revisiting in six months, but do not bet a production system on it yet.