Content Public, documented capabilities onlyClaude architecture
Architecture patterns for building applications with Claude: agentic loops, tool use, MCP, prompt caching, computer use, and extended thinking. These pages cover how to design systems around Claude's public API. They don't cover Claude's internal model architecture, which Anthropic doesn't publish.
Batch processing architecture: the Message Batches APIAsync submit-and-poll instead of request-response, good fits, and idempotent per-item processingBuilding agents with Claude: the agentic loopThe anatomy of one loop iteration, stopping conditions, and approval boundariesClaude and MCP: the Model Context Protocol as an integration patternClient and server, tools vs. resources vs. prompts, local vs. remote serversClaude computer use: architecture and safety boundariesWhen UI-driving beats an API, and why it needs tighter constraintsClaude and retrieval: long context vs. RAG, and citationsWhen to put a corpus directly in context, when retrieval earns its complexity, and what citations addClaude tool use and function calling architectureTool descriptions as an API contract, parallel vs. sequential calls, error designExtended thinking: architecting for Claude's reasoning modeRouting hard requests to deeper reasoning instead of defaulting it everywhereGuardrails architecture: validating and constraining Claude's outputsEnforcement outside the model, fail-open vs. fail-closed, and where a judge model fitsMulti-agent orchestration patterns with ClaudeOrchestrator-worker, context isolation as the real benefit, and coordination failure modesMultimodal architecture: vision and documents with ClaudeWhat vision is reliable at, PDF layout as real signal, and the token cost of an imagePrompt caching architecture for Claude applicationsStructuring prompts so the static part caches, and where it pays offStreaming architecture for Claude applicationsDeltas instead of finished messages, partial JSON, and handling a tool call mid-streamStructured data extraction: schema design and validation architectureSchema design as the interface, validation failure handling, and where extraction fits in a pipeline More patterns will be added. None of them use invented statistics or testimonials.
