Graphcoder: the first step towards hyperscale coding agents

TLDR: We’re launching graphcoder as our first step towards efficient generalised multi-agent scaling. It’s a coding agent that can efficiently run hundreds of agents within a single conversation. Try it out here :)

The amount of compute available to humanity is growing fast. The direction is obvious: each person's work is going to be augmented by more and more agents over time. The question is how do we harness them efficiently.

Right now, the answer is mostly: wastefully. Separate worktrees and/or conversation threads cause excessive context switching. Cloud agents go the other direction and remove the human from the loop entirely; compromising on one or more of speed, agent understanding or human understanding.

None of these are scalable. They're just scaling compute and hoping that the human can keep up, or cutting the human out and hoping the agent gets it right.

What's actually needed is innovation across the stack. We need a UX that keeps the developer in the loop without forcing context switches. We need tooling that allows productive multi-agent collaboration within a single human conversation. We need orchestration that decomposes work and efficiently coordinates/manages agents. And we need infrastructure that allows for greater interaction with the environment, whether that be anything from type checking to computer use. These are different problems, and they need to be solved together.

Graphcoder is our first step along this journey, addressing the UX of multi-agent scaling.Try it out This is a large departure from how coding agents work today, so feedback is very welcome :)

(It is also very beta software right now there will be stuff that’s broken)