Technology RadarTechnology Radar
Adopt

Full deep dive: Claude Code Architecture Breakdown

Anthropic's agentic coding assistant, available as CLI, desktop app, web app, and IDE extensions. Uses a two-layer architecture (agentic harness + LLM backend) with no vector database or RAG — instead relying on direct file reads, keyword search, and large context windows (200K–1M tokens). Leads SWE-bench Verified at 80.8%.

Architecture Overview

Claude Code uses a two-layer design:

  1. Agentic Harness — manages the gather → act → verify loop, tools, permissions, context, and subagents
  2. LLM Backend — Claude Opus 4.6 / Sonnet 4.6 / Haiku 4.5 via Anthropic API (also supports Bedrock, Vertex AI, Microsoft Foundry)

All surfaces (CLI, desktop, web, IDE) share the same harness — no surface-specific logic.

Key Design Decisions

Decision Detail
No RAG / no vector DB Direct file reads via Glob + Grep + Read; context windows large enough to search on demand
Context engineering Auto-compaction at ~95% capacity; CLAUDE.md re-injected losslessly after compaction
Sandbox model Application-level (permissions + hooks), not kernel-level
Config standard CLAUDE.md (Claude-only, supports hierarchy + @import + hooks integration)
Extension model MCP servers, hooks (17 lifecycle events), subagents, custom agents, skills
Token efficiency ~4x higher usage than Codex — trades efficiency for thoroughness

Models

Model Use Case Context Window
Claude Opus 4.6 Complex reasoning, architecture 200K–1M
Claude Sonnet 4.6 Day-to-day coding 200K–1M
Claude Haiku 4.5 Fast tasks, subagent work 200K

Why It's in Adopt

Claude Code's combination of highest benchmark scores (SWE-bench 80.8%), best-in-class reasoning quality, and a mature extension model (MCP, hooks, subagents) makes it the top choice for complex, multi-file coding tasks. The no-RAG approach eliminates setup complexity. The main tradeoff is higher token consumption and CLI-first design (though the web app and IDE extensions have closed the gap).

Key Characteristics

Property Value
Company Anthropic
System Claude Code (CLI + desktop + web + IDE)
Models Claude Opus 4.6 / Sonnet 4.6 / Haiku 4.5
Key innovations No-RAG context engineering, auto-compaction, CLAUDE.md, hooks engine
Open source Partial (Claude Code SDK, MCP protocol)
Sources Overview, Context Engineering, Architecture Breakdown