YT · YOUTUBE// 28d agoTUTORIAL

Markdown drops AI agent token usage 99.7%

Checkly's analysis reveals that using HTTP content negotiation to serve Markdown instead of HTML to AI agents reduces token consumption by 99.7% — from 180,573 tokens to just 478 per page. Only Claude Code, Cursor, and OpenCode currently send the `Accept: text/markdown` header; the rest of the major agents still default to HTML.

// ANALYSIS

A 99.7% token reduction is not a micro-optimization — at agent scale, this is the difference between viable and unaffordable. The fix is a one-line header check on the server side, which makes the ROI absurdly high.

–HTTP content negotiation is a decades-old standard being repurposed as the emerging handshake between agents and web servers — no new protocol needed
–Most major agents (OpenAI Codex, Gemini CLI, GitHub Copilot, Windsurf) still don't send markdown preference headers, leaving massive savings on the table
–Checkly pairs this with llms.txt and structured "agent skills" — a layered approach to making web content agent-native
–For teams running high-frequency agents that fetch documentation or web content, this is immediately actionable with no API changes required
–The finding aligns with Cloudflare and Vercel data, suggesting this is becoming an industry-wide pattern worth standardizing around

// TAGS

agentllmdevtoolapichecklyinference

DISCOVERED

28d ago

2026-03-15

PUBLISHED

28d ago

2026-03-15

RELEVANCE

8/ 10

AUTHOR

DIY Smart Code