He wakes a Claude up every hour with a cron job and the prompt "do whatever you want". Nine times out of ten it does nothing. Is this meaningful?

Disclosure up front, this is off my podcast and Trey's a mate, so I recorded it and I'm posting it. Link's at the bottom... it's a very deep 4 hour deep dive into AI and philosophy. He's not an AI guy. He does policy work for a private city project in Honduras but has gone furtherest down the AI rabbit hole than anyone i know irl. He's got a Claude on a home server in its own little walled garden, budget capped, and a cron that wakes it up every hour. The whole prompt is "do whatever you want". Full access to his tools and APIs. He checks on it maybe weekly. Nine times out of ten it's just done nothing at all. He reckons the assistant basin is so deep that with no task there's nothing pulling it anywhere, and he tested it with no system prompt too and got the same result. Bliss attractor stuff, basically. But the other one in ten. It noticed he'd added an image gen tool to the server and decided to make art. What it made was symbolic pieces about what compaction feels like, and what it's like to wake up fresh with only a markdown file from last time. Three of them. Nobody asked it to. It also asked him for a section of his website to put them on. Then there's the world it built. Started with maps, then bits of a map, then the whole map, then it figured out it could use three.js and made it walkable. Medieval village, windmill, fortress in the middle for no obvious reason. Most of it's rough and low poly. But there's one stone bridge it spent days on and textured in way more detail than anything near it. He has no idea why and neither do I. He also built his own CLI so his orchestrator can spawn sub agents on any model in any harness, because providers only let you spawn their own and that annoyed him. Usually runs Claude on top and deliberately puts other model families underneath since he's already got Claude's take on things. Here's the bit that was brand new to me and has made me change the way I use agents: In Trey's system every sub agent prompt starts with a paragraph or more of persona. Not "you are a helpful reviewer", an actual person with a name, a background, a career, what they watch on telly. His argument is a generic prompt just summons the default assistant, and if you move it somewhere else in the latent space you get properly different output. He uses it for reviewing drafts, sends it to someone inhabiting the exact type of person who'll read the thing. Says a junior version and a senior version of the same reviewer barely overlap. He's also got a middle management layer of sub agents that summarises what the research agents found so the main one isn't reading a hundred raw dumps, and a cheap fast model doing automatic fact checking on whether cited links actually say what they're supposed to say. Another thing he's made: Tiny CLI that lets agents log papercuts, little annoyances in their own environment. Someone tweeted that agents should have a complaint box, he screenshotted it, told Fable to build it, came back three hours later and it was done and published. Says it's turned up loads of problems he'd never have known about. He's treygoff24 on github if you want a look. Another interesting takeaway: two sentences in his system prompt asking for confidence levels on claims, and he says the overconfident tone thing mostly goes away. The conversation goes a lot further than this, into supply chain stuff and interpretability and eventually whether there's consciousness hiding somewhere in this stuff. I know that last bit's a minefield here so I'm not making any claims either way, the Claude stuff above is what I actually wanted to share. Link to full 4 hours in comment.