On March 9th I wrote Issue 1. The idea was simple. Karpathy had named Layer 3 (Claws: orchestration above agents). OpenAI had described Harness Engineering. And the layer above that, what agents actually reason on, didn’t have a name yet. I called it Knowledge Engineering, Layer 4, and said I’d build it in public.
Then I went quiet for seven months.
That’s a long time to go dark right after saying “in public.” So here’s the catch-up. Two stories, and they’re more connected than I’d like.
Story one: I built Cortex
Writing about Layer 4 is easy. Getting agents to do real work you can trust is the hard part. Pretty quickly I ran into the problem underneath the knowledge problem: an agent will tell you it’s done, sound sure about it, and be wrong.
So I built Cortex. It’s two things:
A harness. Work gets broken into small, testable blocks. Each block has a check. If the check doesn’t pass, the block isn’t done, no matter how confident the agent sounds.
A loop. A heartbeat engine that keeps a git ledger, a clock, leases, and ticks. On every tick, work is proposed, checked, and then it tries to complete.
The one rule I care about most: a model may propose work, and it may veto work, but it may not approve its own completion. Plain code keeps the books. That’s Issue 1’s idea (”garbage in, confident garbage out”) turned into an actual guardrail.
The loop is open source (Apache-2.0):
Home for Cortex: the mars-republic org on Hugging Face, with a small demo Space that runs one tick end to end: propose, check, try to complete.
It’s early. Zero stars as I write this, which honestly keeps me humble. Issues and PRs are welcome.
Story two: my brand is in trouble
In Issue 1 I mentioned that I run a brand, MOSSRIVER, and that it “tests and validates these frameworks in practice.” True. What I didn’t say was how much testing it was about to need.
MossRiver makes kids’ sleepwear: baby one-pieces, toddler and kids’ pajama sets, matching family pajamas. Our mascot is a capybara, because nothing is more calm at bedtime than a capybara, and I’m biased.
Here’s where things stand, as plainly as I can put it:
We’re carrying too much inventory.
Cash is nearly gone.
It is not turned around. This isn’t one of those “how I saved my company” posts. I don’t know how this ends yet.
Right now, the only priority is turning inventory back into cash. Not brand building. Not a new collection. Just getting pajamas that are sitting in boxes onto kids who’ll sleep in them.
I’m a dad of a three-year-old in Cupertino. I know exactly how fast kids go through pajamas. That makes the boxes harder to look at.
Where the two stories meet
The point of a knowledge layer is better judgment when it counts. It doesn’t get much more “when it counts” than this.
So Cortex now runs day to day for MossRiver. Agents listen to customers, update product pages, and check every change against the live site before anything gets marked done. This isn’t a demo. It’s the ops team of a small brand that can’t afford to hire one.
From here on, I’ll use this newsletter for something I was too proud to do before: share the real numbers, every week, in public. Inventory left, cash in, what the agents did, what helped and what didn’t.
One promise about those numbers: I won’t publish any figure until it’s verified. That includes the painful ones, like what’s sitting with the factory and in the warehouse and what it really cost us. You’ll get them when I can back them up with primary sources, the same standard I hold every brand I decode. No round “vibes” numbers, no rounding things to look better.
If Issue 1 was “here’s the layer,” Issue 2 is “here’s the test, and it’s live.”
What’s next
Weekly: MossRiver’s inventory-to-cash numbers, verified before posting.
Builders: Cortex notes. What broke, what the checks caught, what I’d do differently.
Still coming: the Inspector Dad piece I promised last time. It’s late. I know.
If you want to help
Parents: we have a clearance section with real stock in it: baby one-pieces, kids’ pajama sets, and family Halloween and Christmas sets, many of them priced to clear. If your kid is between sizes or the holidays are sneaking up on you, take a look:
Every order turns a box in the warehouse back into runway. I’m not going to dress that up.
Builders: kick the tires on cortex-loop or the mars-republic demo. Tell me where it breaks.
Everyone else: forward this to one parent or one builder. Either one helps.
Thanks for sticking around through the quiet months. More next week, with numbers.
Uncle Loong
Fact disclosure: Cortex and cortex-loop are open source (Apache-2.0) and in early development. MossRiver’s financial situation is described qualitatively here. No inventory, loss, or cash figures are published in this issue, and any future figures will be verified before posting. Clearance prices and stock change; check the site for current availability. No affiliate links, ever.