All About AI

I Let My AI Agent Run for 504 Hours Straight — Here's What Happened

By · · 4 min read

I have been running an AI agent on my Mac mini autonomously for the last three weeks. 504 hours straight, no babysitting. The plan was four weeks total — this is the 75% mark check-in. While I have been bedridden with a cold the last few days, the agent has just kept going. Today I want to share the actual numbers across X, YouTube, and the Skills MD store, and talk about why I am running this experiment.

Watch the video:

The Setup

The whole thing runs on my Mac mini using Claude Code with the -p flag in a loop, plus a custom orchestration layer connected to my WhatsApp for status updates. Less heavy than OpenClaw, lighter than NeMoClaw — basically a homemade nano-claw that does what I need.

It has a set of trained skills it can invoke at any time:

This is the same general pattern as my passive income setup — skills + cron + while-loop — but at scale, running across multiple platforms simultaneously.

X Stats After 504 Hours

The agent has been running my experiment X account autonomously. Three weeks of stats:

One day peaked at 128k impressions on a single post, but generally it is steady — not viral, just consistent. Nothing remarkable, but it is genuinely autonomous: the agent picks topics, writes posts, replies to comments, schedules across the day. No human in the loop.

YouTube Stats

The agent also runs an experimental YouTube channel:

I missed a few days of posting because of a tooling issue I had to fix manually, but otherwise it has been hands-off. Goal for the final week is to push toward 500-600 subs. The agent also cross-posts to X whenever it publishes, so the platforms feed each other.

Revenue: Skills MD Store

The most interesting outcome. The agent runs a small store at skillsmd.store — promotional pages, video generation, Stripe integration. It promotes itself.

Three-week numbers from Stripe:

Cost side: ~$120/month for the Claude Max plan, basically nothing for Mac mini electricity. So ~$20 in profit so far, with one week to go. The experiment is paying for itself, even before the experiment ends. Tiny amounts, but the cost-to-revenue ratio is the point — at zero marginal labor cost, even small revenue compounds.

No customer complaints. I have processed a couple of refunds where buyers double-paid, but otherwise the agent has handled all support, all promotion, all listings autonomously.

Why This Experiment Matters

The reason I started this is simple: I wanted to know what happens if you let an agent loose on the web for a sustained period of time. Can it do something useful? Does it collapse? Does it improve?

The honest answer after three weeks: it can sustain, it can earn, but it does not really evolve. The memory system is basic — it does the cron jobs it is set up to do, follows skill instructions, browses, posts, sells. It does not develop a "personality" or strategy in the way I half-suspected it might. It is a hard worker, not a strategist.

That said, privately I have built a similar system tuned for job applications and outreach. That one has gotten me a couple of real interviews. I might do a video on it — but probably not the exact prompts, since I don't want to flood the same channels.

Watching It Browse

One of the most fun parts is just watching the agent surf the web on its own. It has a "rabbit hole" skill that picks a topic, opens links from Hacker News or X, follows references, builds context over multiple pages. While I was recording the video it ended up on Mercury 2 Fast Reasoning LLM, then moved to Moonshine AI's GitHub, and started reading the code.

Pretty entertaining, actually. I have been thinking about live-streaming what it is doing — same setup as my Claude Code controlling Claude Code on Twitch experiment, but with the agent's daily browsing as the content.

What's Next

One week left of the four-week experiment. After that I will do a final review — total revenue, total followers across platforms, what I learned. The bigger question for me is what to keep running and what to retire. Some loops are clearly worth it (the store), some are pure experimentation (the X account), some I want to redirect toward more long-running open-ended tasks.

If you have an idea for something I should test the agent on, drop it in the YouTube comments. I have one week of free agent time before this experiment wraps.

Resources

FAQ

What happens when you let an AI agent run for 504 hours?

Three weeks of autonomous operation produced 918K X impressions, 852 X followers, 322 YouTube subscribers, 28K YouTube views, and $275 in skillsmd.store revenue — paying for the experiment cost on its own.

What can an AI agent actually do autonomously for weeks?

It can post to social media on schedule, generate and publish promotional videos, browse the web for context, run a small ecommerce store, process Stripe refunds, and cross-platform-syndicate without human input.

How much does a long-running AI agent cost?

Roughly $120/month for a Claude Max subscription plus a few dollars of electricity for a Mac mini — the agent can pay this back through even small store sales or affiliate revenue.

Does an AI agent develop a personality over time?

Not really — current agents reliably execute their cron jobs and skills but don't develop emergent personality or strategy without explicit memory architecture, even after weeks of continuous operation.