Day 10 building a codebase agent.
Learned: judging work is harder than doing it.
Built a loop where the AI checks its own work.
Small model → never approved anything, nitpicked forever. Bigger model, same code → approved first try.
The checker has to be the smart one.
Hey man, random but been seeing you get slammed on X recently, just wanna say I appreciate the hard work you do
If it wasn't for OpenClaw we wouldn't have the massive progress across these kinds of personal agent harnesses
OpenClaw was the first agent harness I used, and now I get to enjoy other harnesses too. Thank you Legend, keep it up
Here is the thing, I made something like this but the context is different. It works well until maintaining that knowledge in the workspace and the agent. After some time, it again requires maintaining memory and sessions, and every time I get to experience this.
Doing my first ever experiments with a personal, entirely agent-managed Karpathy-style wiki
X, Discord, Gmail are all being ingested into it every few hours
This is the knowledge base that will serve as the environment for all of my future loops
If I were starting a new company today, I'd start a small Iranian restaurant just serving “Chelo Khoresh” in Germany because they and the world don’t know those special tastes.
If I was starting a new company today, I'd start a small pie shop on the corner somewhere. Counter service with some stools, incandescent lighting, just a few pie choices in a rotating display, and a mug for coffee. Probably call it "A piece of pie and a coffee, please". One flat price. Coffee and a slice of pie included. Don't want the coffee? Don't drink it. Don't want the pie? Don't eat it.
Today we're announcing Live Studio, a brand new livestreaming command center on X
X is where everything is happening now. So we're launching the best tools for pro streamers to go live, connect with their followers & manage their streams
Check it out on 𝕏.com in Creator Studio
Day 9 building a codebase agent from scratch.
Layer 1 was one-shot retrieval.
Layer 2 is about workflows: breaking big tasks into smaller model calls.
Today: restructured the project, finished the Layer 1 README, studied the foundation, and started implementation.
Slow progress is still progress.
Most search scores the question and the answer separately, then compares.
A reranker reads them together and asks one thing: "does this actually answer what they asked?"
Small shift. Huge difference in accuracy.
A lesson from today's code that turned out to be about people:
You don't always need more opinions in the room.
Sometimes you need one person who actually understands the question.