Claude Code Lapsed. The Reactivate Button Is Right There.

I opened a different coding agent last week and sat there for actual minutes, stuck on something that costs me zero seconds in Claude Code. Not a bug. Not a missing feature. My hands just didn’t know what to do without the specific shape of harness I’m used to.

It stopped me more than it should have. I’ve written two posts now about how the model is the commodity and the harness is where the real engineering lives, how the loop and the memory dict and the tool schemas are the actual work, not the API call underneath them. I believed every word of that while writing it, and I still mostly do. I just hadn’t noticed I’d only ever tested the idea inside one harness, built by people who are extremely good at building harnesses.

It’s a narrower kind of expertise than it feels like from the inside, probably. I know exactly how Claude Code handles a permission prompt, how it threads context across a long session, how it decides when to compact and what it keeps. I have real opinions about all of that. The thing I don’t actually know is which of those opinions are general truths about good agent design and which ones are just how Claude Code happens to do it, dressed up as a principle because I’ve never had to look at the alternative.

Same mistake the last two posts were about, honestly, just aimed at myself this time. Knowing one city’s subway by heart isn’t the same as understanding transit. I won’t know which of my rules are load-bearing until I take them somewhere they might not hold, and I don’t know yet how many will hold.

Claude Code lapsed. I can always go reactivate it. I know where the button is. I am just doing this independently for a while. Model-agnostic, cost-efficient. I keep saying those two words like a bit. The restaurant agents are where I learned how expensive a product can get once you are calling a bunch of external models and the tokens start stacking. OpenRouter, Baseten, picking where a call actually runs instead of treating one vendor as the whole stack.

So that’s what the next few posts are. Not a review roundup, not “which agent is best.” I’m going to sit inside a handful of other harnesses long enough to actually build something real in each one, and check my Claude Code instincts against them one at a time. Same questions every time: how it handles memory once a conversation outgrows the context window, how tool dispatch actually works underneath, what it does the moment something fails halfway through a task. Where my instinct holds up, I’ll have an actual principle instead of a habit. Where it doesn’t, that’s the more interesting post.

Some of what I’m shipping next is agent-native deployments. I’ll post the build logs and the workflows as they land, over the next few weeks. I am done mistaking fluency in one harness for understanding harnesses.