Researchers Reverse-Engineered Claude Code Architecture

A team of researchers dug into Claude Code's TypeScript source and published a detailed architectural breakdown of how Anthropic's agentic coding system actually works under the hood. The core loop is deceptively simple - call the model, run tools, repeat - but the surrounding infrastructure is where the real complexity lives. They identified a seven-mode permission framework with ML-based classification, a five-layer context management pipeline, and four extensibility mechanisms including MCP and subagent delegation with isolated worktrees. The paper compares Claude Code against OpenClaw, an open-source alternative, and finds that deployment context drives most design divergence between the two. As more teams ship agentic developer tools, this kind of teardown becomes genuinely useful reference material for anyone building in the same space.