Your agent spends 55,000 tokens before it reads your prompt11 min readGitHub's MCP server costs 55k tokens before the first message. How AGENTS.md, Skills, and MCP split the agent context problem and converged on lazy loading.4 June 2026·11 min read