Background Advisory Ventures Dispatches Contact
Autoresearch Runs While I'm Away
All dispatches

Autoresearch Runs While I'm Away

First shared in my AI club, now Frontier Table, April 8, 2026.

A version of Autoresearch is now in every project on my desktop. When I know I’ll walk away for the next hour or so, I just type “AR” and that project improves constantly, confined to existing functionality, until I get back to it. Built that right quick in Claude Code.

The idea is Andrej Karpathy’s. He published autoresearch on March 6 as a small repository for “AI agents running research on single-GPU nanochat training automatically.” The loop is the whole thing. In his words, the agent “modifies the code, trains for 5 minutes, checks if the result improved, keeps or discards, and repeats.” His version trains a model overnight — “You wake up in the morning to a log of experiments and (hopefully) a better model.” Mine is for ordinary software and an ordinary hour away from the desk, but the shape is the same: change one thing, measure, keep or throw away, go again.

The fence matters as much as the loop. Confined to existing functionality means it can’t wander off and invent features while nobody’s watching. That limit keeps the agent focused while I’m away.

On other projects — I’m skeptical of trying them, and happy to be convinced otherwise. My version of OpenClaw is now so customized, and there are so many people working on the project, that I just presume any innovation will land in OpenClaw in seven more days. So I wait and preserve my adjustments. For a sense of scale, by late February OpenClaw had passed the Linux kernel on GitHub stars, “with 224K+ stars” against the kernel’s 218K. That is a lot of people carrying features toward you for free.

I also think security will end up being tighter in OpenClaw, because it’s that much harder to outrun that project’s data or contributors. The big vendors seem to think so too. NVIDIA announced NemoClaw on March 16, a stack for OpenClaw that installs its Nemotron models and a new OpenShell runtime “in a single command” and adds “an isolated sandbox that adds data privacy and security to autonomous agents.”

I love to show off these technologies and put them through their paces. I’m not a teacher or a presenter, though.

Ragu is testing things like NemoClaw. Will report.

Where this started

Most of these dispatches start as posts in the Frontier Table, a group of about 100 builders, founders and operators trading what's actually working with AI.

Robertson Price

Robertson Price

Serial entrepreneur who has built and exited multiple internet companies over 25 years — from search (iWon.com, $750M acquisition) to content networks (32M monthly visitors) to e-commerce (Rebates.com). He now builds enterprise AI infrastructure at Ragu.AI.