Forking Reality: Coding Agents in Branchable VMs

Demonstration of LLM agents spawning parallel VM branches, each compiling and testing code autonomously, with real‑time, survival‑of‑the‑fittest evaluation process, feedback loop.

Overview

What happens when coding agents don’t just suggest code, but spin off entire parallel attempts? Each one free to compile, test, and crash on its own path.

We’re going to walkthrough how I wired up a swarm of LLM-based agents to collaborate like YC interns on Celsius.

This is freaking beam search for software: dozens of competing code branches running in parallel, survival-of-the-fittest style, until something actually works.

No slides, no pitches. Just raw orchestration, messy agent chatter, and live experiments of coding intelligence at scale.

Tech stack