Claude Code sparks context-window debate

This is a useful critique of the whole agentic coding stack, not just a shot at Claude Code. The better these tools get at carrying context for you, the more important it becomes to understand how much context they quietly consume.

• –Anthropic positions Claude Code as an agentic coding tool that works across terminal, IDE, desktop, Slack, and web, so heavy built-in context is part of the product, not an accident
• –Anthropic’s docs confirm Claude Code loads `CLAUDE.md` instructions and memory at session start, which helps continuity but also adds to baseline prompt overhead
• –The video’s core argument is that large system prompts and MCP/tool context can crowd out the usable working window for the code and reasoning you actually care about
• –This matters most in long coding sessions where repo instructions, tool output, logs, and file context all compete for the same finite token budget
• –Expect context efficiency to become a sharper competitive axis across AI coding agents, alongside model quality and tool integration