This is f*cking gold..
A RULE WON 11 OF 15 TRADES AND TURNED $10,000 INTO $10,770. ONE BUY DID SOMETHING ELSE.
it's vn. py, 46k stars, the open-source quant platform out of china.
five years on amd, and the rule does one thing: buy the dip, sell the bounce.
it was right 11 times.
it paid 2.90% in fees to be right.
and its drawdown still hit -57%.
honestly a 73% win rate looks great in a screenshot. the final balance doesn't care.
that one buy is on screen in the clip.
Stop tuning prompts when your AI agent breaks.
The model is the part you can't change. The harness is the code around it, and you can change that.
An Oracle developer advocate explains it in about an hour.
These are the moments that matter.
13:43 - the seven layers of a harness
16:26 - files versus databases for agent memory
18:30 - work trees for parallel agents
28:48 - context rot as the window fills
37:24 - toolbox pattern for many tools
Free, and better than many paid agent courses.
Watch the video first, then read the article below on how Stripe, Uber and Spotify build their agents.
We are not f*cking ready for this
I put 500 Claude Code agents to work in parallel and gave them one boss: a verifier that kills any claim it can't trace back to a source.
My live dashboard mid-run:
- 500 agents researching and writing at once.
- Every output verified against the source in ~60 ms.
- 6 passes over 200 company profiles → 0 errors left.
- Every verified note goes straight into my Obsidian vault.
The trick is the split: one model does the work, a cheaper one decides if it's actually true.
Nothing gets saved until it checks out.
An analyst team would bill weeks for this. I got it in one run.
Full breakdown below. Bookmark it.