The Shift from Human to Agent Onboarding

As AI agents increasingly influence developer tool selection, the traditional model of human-centric documentation is insufficient. The author observed a significant spike in inbound traffic to their library, c15t, directly attributed to LLMs recommending it. The core insight is that developer experience (DX) primitives have evolved into agent primitives. Because agents often struggle to browse complex websites, developers must treat documentation as a data-delivery problem.

Optimizing for Agent Consumption

• Curate, Don't Generate: While tools can automate documentation pipelines, hand-write your llms.txt. Forty lines of high-quality, human-curated context outperform a thousand lines of automated noise.
• Provide Markdown Alternatives: HTML is expensive for agents to parse. Serve markdown versions of your pages via a .md suffix, content negotiation headers, or a ?mode=agent query parameter. Always signal the availability of these versions in your site header.
• Ship with the Code: The most critical realization is that coding agents rarely visit your website; they read your repository and node_modules. By bundling an AGENTS.md file and markdown documentation directly into your package, you provide agents with immediate, accurate context. This approach can reduce token usage by nearly 50% compared to forcing an agent to search the web for documentation.

Future-Proofing the Pipeline

Because the landscape of agent-ready web standards changes weekly, avoid seeking perfection. Instead, focus on building a framework-neutral pipeline that transforms your source (e.g., MDX) into agent-optimized files. As web-based Model Context Protocol (MCP) evolves, ensure your documentation infrastructure can expose specific tools—such as search, page retrieval, and Q&A endpoints—directly to agents. Use testing tools like Aura AI to benchmark your site's agent-readiness, but accept that scores will fluctuate as the underlying models and standards evolve.