Ending the Agent Industry’s Wheel‑Reinventing: ADPS Launches the First Global Agent Design Language

Industry Pain Points

Agentic AI ecosystems are fragmented: dozens of frameworks (OpenAI, Anthropic, Google, Doubao, Kuzi, Cursor, LangGraph, MCP) exist without a common design language. Teams concentrate on rapid tool calls, prompt stitching, and demo construction, while core decisions—topology selection, memory layering, reflection‑governance, and cost‑performance trade‑offs—lack standardized evaluation, leading to repeated reinvention and isolated engineering knowledge.

Double‑Axis Framework and Design Patterns

ADPS introduces a two‑dimensional framework:

Cognitive Function Axis (What) – defines seven foundational capabilities that consume resources.

Topology Execution Axis (How) – defines six structural forms for resource consumption.

Cognitive functions :

Perception – context triage, semantic compression, multimodal fusion, progressive discovery.

Memory – layered memory (working, short‑term, long‑term vector store, program skill), RAG pipelines, progress tracking, failure logs.

Reasoning – chain‑of‑thought, complexity routing, parallel exploration, iterative hypothesis testing.

Action – tool‑call sandwich, plan execution, environment interaction, MCP‑based routing.

Reflection – generate‑evaluate loop, self‑healing, experience replay.

Collaboration – multi‑agent delegation, hand‑off chain, fan‑out aggregation, adversarial review.

Governance – approval gates, progressive commitment, explosion‑radius control, observability harness.

Topology forms (derived from Anthropic, LangGraph, Google ADK, OpenAI Agent SDK): Chain, Route, Parallel, Orchestrate, Loop, Hierarchy.

Combining the axes yields a 7×6 matrix (42 theoretical coordinates). ADPS distills 28 of these into production‑validated design patterns; the remaining slots are left blank for future extension, following the GoF approach.

Core Deliverables

Standardized Pattern Catalog – 28 patterns, each with problem definition, architecture diagram, applicable scenarios, failure modes, code snippets, and cost trade‑offs. Code is hosted at github.com/huangjia2019 and will migrate to an ADPS organization.

Agent Selection Methodology – a six‑step workflow plus five core design laws that provide a common evaluation rubric for architecture choices.

Open‑Source Production‑Grade Harness – reference implementation that enforces budgeting, constraints, routing, auditing, and safe‑stop capabilities, demonstrating how multiple patterns combine to achieve cost‑controlled, risk‑mitigated production agents.

References

ArXiv paper “A Two‑Dimensional Framework for AI Agent Design Patterns” and Manning monograph “Designing AI Agents”. Official site (plain text): https://adpsagent.com/zh.