Yeah, probably pretty simple compared to the methods we've publicly discussed for months before this publication.

Here's the last time we showed our demo on HN: https://news.ycombinator.com/item?id=45132898

We'll actually be presenting on this tomorrow at 9am PST https://calendar.app.google/3soCpuHupRr96UaF8

Besides ReAct, we use AG2's 2-agent pattern with Code Writer and Code Executor in the DockerCommandLineCodeExecutor

Also, using hardware monitors and LLM-as-a-Judge to assess task completion.

It's how we've built nearly 1K Docker images for arXiv papers over the last couple months: https://hub.docker.com/u/remyxai

And how we'll support computational reproducibility by linking Docker images to the arXiv paper publications: https://github.com/arXiv/arxiv-browse/pull/908