$8,679 IN, $15,578 OUT. A DEV SOLD 18% OF THEIR OWN COIN, GROK ASKED TO SHORT IT 10 MINUTES LATER, AND JEV SAID YES BEFORE I'D FINISHED READING THE CARD.
the clip is grok bot + jev. grok proposes, jev judges, nothing moves without both.
1,876 pairs scanned → 4 pass → short noodl → jev 0.94 → approved
"dev sold 18% 10m ago · LP $96k · thin" - the noodl card
my favourite part is the log in the corner. it reads like a receipt. flint, 0.14, blocked. noodl, 0.94, approved. no adjectives.
0.94 was the biggest yes of the whole session. it went to a short.
+79.5% on the session.
the tallest green step on the curve is on screen in the clip.
Jev is TypeSafe's System One model, from the company led by GPT-4 co-author Diogo Almeida.
bookmark & watch today ↓
yo, this is f***ing wild
Grok Bot engineer Lauren Tan revealed her agent formula:
open Grok Bot and use skill-to-automation.prompt to move from manual corrections to verified, repeatable execution
Paste the operating contract:
role: one bounded responsibility
outputs: artifact + checks + blockers
verify: observable acceptance criteria
escalate_on:
- missing_access
- unclear_scope
→ keep preferences + skills persistent
→ keep evidence + progress local to the task
Test the revision against the current skill:
current_skill vs revised_skill
same tasks, model and environment
record outcomes / failed_checks / human_corrections
keep the update only if observed execution improves
The automation gate:
repeated_success + clear_trigger
> define allowed actions and stop conditions before scheduling
> audit polling frequency and duplicate jobs
missing access or unclear scope keeps the task supervised
Bookmark this to give your agents a clear path to autonomy