Anthropic gave the failure mode I kept hitting with Claude Code a name: agentic technical debt!

Anthropic calls a common AI-coding failure mode “agentic technical debt”: repeated sessions can drift from the original design.

Tools like Claude Code, Cursor, and Cline may duplicate implementations and slowly push the codebase toward inconsistent architecture.

The article argues the fix is not more memory or better prompts, but written specs, clear scope boundaries, and a disciplined session workflow.

ADRs and persistent docs like CLAUDE.md help constrain the agent’s choices over time and keep it aligned with the intended architecture.