AI Pulse · Under the Radar

Boris Cherny, the guy behind Claude Code, posted something worth paying attention to if you’ve been hearing “AI agents can just run your business now” and wondering how true that is. He set up a dedicated Slack channel called proj-claude-maintains-apps, where Claude proposes and ships its own maintenance pull requests on his apps. No human writing the code, just reviewing and approving what the model comes up with.

Here’s the part that matters: it took a few days of tuning before this worked reliably. Not minutes, not one prompt. He had to adjust how the system was set up before Claude could be trusted to maintain real, running software without someone hovering over every change. And this is from the person who built the tool. If the creator needed days to get it running smoothly, that tells you something about what “unattended AI” actually requires in practice.

Multiple X threads picked this up within a day or two, mostly agreeing it’s a useful data point on where autonomous coding agents really stand right now, as discussed in the original thread.

For business owners, the takeaway isn’t “AI can’t be trusted yet.” It’s that autonomous systems need a setup period before you hand them real responsibility. If you’re exploring AI agents for internal tools, customer workflows, or ops, budget time for tuning and a review loop, not a switch you flip on day one. That’s exactly the kind of staged rollout we think about when we build client systems, this is the kind of thing we build into an AI command centre so the tuning happens in a controlled space before anything touches production. Go in expecting a runway, not instant autopilot.