Skip to main content

Real World Systems

How the most widely-used AI systems are actually built. Each breakdown maps the system's architecture to the patterns in this catalog — showing which patterns they use, how they combine them, and where they diverge.

These are not speculative designs. Every detail is sourced from official documentation, open source code, published research, reverse engineering, and system cards.


The systems

SystemBuilderTypeCore Pattern
Claude CodeAnthropicCoding agent (CLI)ReAct loop + sub-agents
ChatGPTOpenAIConversational AIRouter + tool orchestration
CursorCursor Inc.AI code editorMulti-model routing + RAG
DevinCognition LabsAutonomous coding agentPlan-and-execute + full VM
GitHub CopilotGitHub / MicrosoftCoding assistantFIM completion + RAG
PerplexityPerplexity AIAnswer engineRAG pipeline + citation
OpenClawPeter SteinbergerAutonomous agent platformChannel adapters + sandboxed tools

Patterns in the wild

Every system in this section combines multiple patterns from the catalog. Common combinations:

Coding agents (Claude Code, Devin, Cursor Agent) tend to use ReAct or Plan-and-Execute for reasoning, a Tool Router for function calling, and File-Based Memory for persistence.

Conversational systems (ChatGPT) lean on Router patterns for tool selection, layered Guardrails for safety, and multi-tier memory systems.

Search-augmented systems (Perplexity) are built around RAG with Citation and multi-stage ranking pipelines.

IDE-integrated tools (Cursor, Copilot) combine RAG for codebase context with speculative decoding and model routing for latency optimization.

Messaging-first agents (OpenClaw) use a Router pattern for channel normalization, File-Based Memory with hybrid semantic search, and session-based trust boundaries with Human-in-the-Loop pairing.