Introducing Raysurfer

Last week I left Modern Realty (YC S24) to work on founding a different AI startup. To those that helped me along the way build that company, thank you so much.

I left to fix something that has been bugging me about the way people currently run tool-using LLM agents.

LLM agents that use tools today are typically architected to take very similar inputs and regenerate very similar tool call outputs in a loop, while dragging every tool input and output along for the entire duration of execution.

This is an issue for both speed and consistency.

Speed: Even with KV caching, there is an approximately linear latency increase for time-to-first-token response with an increase in input length. Much time is spent waiting for LLM outputs.

Consistency: The fraction of context which is relevant for the next step decreases per step; context rot.

Opus 4.5 might run today at 10-30s per iteration.

But it doesn't need to be like this.

The success of Claude Code and the rise of claude-agent-sdk shows us that the next wave of LLM agents can run code scripts that they create/find on the fly. Instead of running 10 tools and piping outputs from one to the next, if the LLM agent knows a sequence of tools will need to be called serially, it can make a script to call those 4 APIs, attaching each of the output variables into the next API calls, then assess where it's at, then make a script to run the next 6 APIs.

The future of context management is print().

Presuming that this is the future, my next step is to build more tooling to further improve claude-agent-sdk.

An early version is available here at raysurfer.com.

If you'd like to chat about it, come talk to me in this Discord channel: