Agent System Architecture from Scratch

This is an agent system architecture tutorial for AI product managers, AI application designers, and engineers. It does not begin with a specific framework or repeat the internal implementation of an agent harness. It studies how agent components should be organized into a complete system.

What This Project Solves

Agent Internals
    = Loop / Tools / Session / Sandbox / Permissions
    -> covered in depth by learn-dsh

Agent System Architecture
    = Workflow / Planning / Context Flow / Memory Topology
      / Multi-Agent / Communication / Orchestration / Protocols
    -> covered by this project

The project focuses on these questions:

• When should a system use a workflow, single agent, or multi-agent architecture?
• What problems do planners, supervisors, handoffs, and evaluators solve?
• How should context and memory flow between agents or remain isolated?
• How do multiple agents communicate, divide work, schedule tasks, and terminate?
• How do graphs, state machines, events, and protocols organize complex systems?
• How do we evaluate, observe, and improve agent-system reliability?

Complexity-Driven Learning Path

LLM
  -> Augmented LLM
  -> Workflow
  -> Single Agent
  -> Planning / Context / Memory
  -> Multiple Agents
  -> Orchestrated Agent System
  -> Distributed Agent System

The core path contains 18 modules:

1. Agent Architecture Overview
2. Workflow Patterns
3. Single-Agent Reasoning
4. Planning & Search
5. Context Architecture
6. Memory Architecture
7. Multi-Agent Fundamentals
8. Multi-Agent Patterns
9. Hierarchical Systems
10. Agent Communication
11. Agent Protocols
12. Orchestration Architecture
13. Coordination & Scheduling
14. Human & Governance
15. Reliability
16. Evaluation & Observability
17. Architecture Comparison
18. Build a Real Agent System

Theory, Practice, and Product Perspective

Every module combines architectural theory, diagrams, a minimal implementation, failure cases, use and avoidance conditions, trade-offs, and framework mappings.

Every core article also contains a dedicated product perspective explaining:

• which user or business problem the architecture addresses;
• which plans, progress, results, and errors users can see;
• the agent's autonomy, data, and permission boundaries;
• how the system degrades, recovers, or hands off to a human;
• which product metrics show whether the added complexity is worthwhile.

The product perspective is a cross-cutting section within architecture articles. It does not replace the agent system architecture learning path.

Current Content

• From AI Feature to Agent Product (Chinese draft): a product-oriented supplement to the overview.
• Chat, Workflow, and Agent (Chinese draft): a supplement about interaction interfaces and execution architecture.
• Basic Agent Loop: an optional technical warm-up that will later move to the appendices.

The first formal article in the main path will be Agent System Architecture Map.

Content Maintenance Rule

Chinese is the canonical content source, and English maintains a complete mirror. The main tutorial is organized by stable architectural concepts; framework details, protocol versions, and advanced patterns live in separate extension areas.