Bound the goal
One employee, one field, one effective date; no retroactive payroll change.
Agent Design Patterns Society
A public pattern language for production AI agent systems.
ADPS is an independent research community that publishes technical references for building production AI agent systems. We bring together software design research, open-source harness engineering, and attributed enterprise case reports to turn common architectural problems into reusable pattern specifications that engineering teams can adopt.
Dual-axis framework
Pattern coordinates for agent systems
Seven cognitive functions map system responsibilities; six execution topologies describe runtime structure. Cross-cutting engineering planes and the lifecycle add system-wide constraints and temporal boundaries.
27 matrix patterns · 3 cross-cutting engineering planes · 1 agent lifecycle
Start with one business change
How an agent turns one sentence into an accepted result
Example request: “Change employee E-1842's monthly transport allowance from 800 to 1000, effective next month.” A model can understand the sentence. A production system must also establish the goal, facts, authority, action, and result.
Minimum modelling unit
One business change that can be decided, approved, executed, accepted, and compensated independently.
Obtain facts
Current value comes from the HR API; policy comes from a versioned knowledge source.
Form the plan
Read, validate, prepare, approve, commit, and read back, each with a completion condition.
Commit intent
Tool, arguments, and resource become an Intent; review covers that exact content and one employee.
Accept the result
The transaction receipt and after-read must agree before the run can claim completion.
Granularity rule
Split again when decision, approval, execution, acceptance, or compensation requires a different owner or rule. The unit is small enough when all five can close independently.
Vision
We want critical design decisions in production agent systems to have a shared language and engineering evidence that can be inspected. ADPS advances that work through four connected sets of public technical references:
- Pattern white papers and module overviews define problems, mechanics, applicability, failure modes, and verification criteria.
- Engineering Blue Books record constraints, failures, architecture decisions, and transfer boundaries in attributed systems.
- Workshops and topics preserve field discussion and synthesize methods that cross several modules.
- Open reference assets include papers, selection methods, implementations, and engineering concepts.
The framework
The two-axis framework — cognitive function × execution topology — was introduced in A Two-Dimensional Framework for AI Agent Design Patterns (Huang & Zhou, arXiv:2605.13850) and developed further in the Manning book Designing AI Agents. ADPS now stewards the catalog as a community asset.
Workshops and topics
Perception, Memory, Action, Reflection, and Governance have completed their first workshops. Workshop records preserve field questions and unresolved differences. Topic pages synthesize engineering issues that cross several modules.
WORKSHOPSDiscussion records, participants, and open questions from five modules
TOPICSCross-module engineering methods drawn from patterns and cases
- AI-driven software engineeringSpecifications, context, execution, verification, and evolution.
- Observability-driven agent evolutionRuntime trajectories, component versions, change evidence, and rollback.
- Agent evals and testingDeterministic tests, trajectory evaluation, external acceptance, and release gates.
Community
ADPS was initiated by Jia Huang, Bingsheng Ru, Willem Jiang, and Bo Liang.
Jia Huang (黄佳): agent architecture, pattern language, and engineering education. Profile PDF
AI researcher at A*STAR Singapore and author of Manning's Designing AI Agents.
Bingsheng Ru (茹炳晟): software engineering, agent systems, and R&D productivity.
Industry Research Consultant at Tencent Research Institute and Vice Chair of the CCF TF Multi-Agent Systems Engineering SIG.
Willem Jiang (姜宁): open-source infrastructure, system integration, and workflows.
Long-term contributor to the Apache Software Foundation and developer of DeerFlow.
Bo Liang (梁博): enterprise SaaS, financial services, and production systems.
Vice President of Shanghai Dongfang Yiteng Technology, working on financial-services SaaS and production execution agents.
See Team and contributors for the advisor, expert group, and White Paper and Blue Book contributors.
Blue Book practice
Three reports document concrete architecture choices in enterprise SaaS, GIS data publication, and MBSE / SysML research.
- Dongfang Yiteng Execution Agent — A task DAG fixes execution order, a mechanical state plane owns parameter provenance, and checkpoints support recovery.
- Xuanxu GIS Data-Publication Agent — Design-time reasoning becomes a versioned asset, runtime follows a deterministic pipeline, and an external probe checks the published result.
- Natural Language to SysML — Typed intermediate representations carry intent; scope resolution and structural write-back gates stop ambiguity before persistence; each operation returns a transaction receipt.
Assets
- arXiv:2605.13850 — the position paper introducing the two-axis framework.
- github.com/huangjia2019/agent-design-patterns — runnable catalog of agent design patterns.
- github.com/huangjia2019/designing-ai-agents — companion code to Designing AI Agents (Manning).
- Manning · Designing AI Agents — the design-pattern catalogue for production AI agents.
Reach out
ADPS is open to engineers and organizations with field experience building production-grade agent systems — at scale, in regulated industries, or under hard reliability constraints. Membership is by contribution, not application.
Fill the three fields below and we'll get in touch through WeChat. Nickname is fine — no real name required.
Prefer GitHub? Open a discussion or issue on agent-design-patterns.