Context engineering is sleeping on the humble hyperlink
6 months ago
- #Hyperlinks
- #Context Engineering
- #LLMs
- Context Engineering for LLMs involves managing the tension between providing all valuable context and only when it's relevant.
- Popular solutions include Retrieval Augmented Generation (RAG), Subagents, and Tools for proactive information requests.
- Hyperlinks are an underutilized technique in agents, allowing dynamic loading of relevant context similar to human web browsing.
- HATEOAS (Hypertext as the Engine of Application State) principles can be adapted for LLMs, enabling self-describing APIs.
- Implementing a link-based context system requires a tool that accepts URIs and an entrypoint to bring URIs into context.
- Links are efficient, flexible, and token-efficient, providing just-in-time context to mitigate issues like context rot and recency bias.
- MCP Resources can enable universal link-based context engineering, but current clients lack full integration for model consumption.
- A proposed solution is a `read_resources` tool for MCP-enabled agents to access linked content across servers and web URLs.
- Hyperlinks offer a simple and powerful alternative to building multiple specialized tools for agent context management.