Sudo Sunday #1: Auto Research

We’ve been curious what happens when you hand a coding agent an unsolved problem and leave it alone for a while.

Someone built a tool for exactly that. Auto-research-agent lets you pick a GitHub repo with an auto-research.md file, press play, and let Codex or Claude Code chew on it with whatever tokens you didn’t burn this week. You can run several agents at once, each on its own or passing checkpoints back and forth.

Its early benchmarks say the sharing doesn’t really speed anything up. That’s a little surprising, and it’s where we want to start.

Stuff we want to talk about over coffee

• Is running five agents for an hour better than running one for five hours? Or is it just a bigger bill?
• When agents share what they’ve learned, should they pass along the whole checkpoint, a summary, or nothing?
• An agent says it solved something. How do you know it actually did, and who’s reviewing all those PRs?
• Which problems are worth handing off so you can walk away?

If you’ve got an agent setup you like, or a run that went somewhere deeply weird, bring it and we’ll throw it up on the screen.

Sunday morning at our office on 2nd St. Coffee’s on us. sudo in.