I run the AI Profit Boardroom.
I build the Agent OS.
I made Goldie Bench so you could see which AI model actually ships.
Short version: I test frontier AI models on real builds every day. Same prompts, one shot, finished things out. This page is where I put my number on each one — honest, public, your call.
Who I am
I'm Julian Goldie. Started in SEO over a decade ago — built and sold one agency, then started teaching what was actually working on YouTube. That grew to 400k+ subscribers and a daily habit of showing my workflows on camera.
Mid-2025 I switched the focus from "SEO tactics" to "AI agents that do the SEO for you" — and started building the Agent Operating System: a Mac-native dashboard that runs a crew of AI agents (Hermes for code, Claude for writing, GLM for long-context research, Kimi for builds) under one shared memory and one set of prompts.
That stack now powers the AI Profit Boardroom — 3,600+ founders inside a paid community ($59/mo), shipping 258+ documented wins across 38 countries, $100k+/mo in community MRR.
What I'm actually doing every day
When a new frontier AI model drops, I do three things, in order:
- Wire it into Agent OS. Make it dispatchable from the same kanban my AI crew uses.
- Run my fixed prompt set through it. One shot each. No iteration. No "here's a hint." Whatever comes out of the model on the first try is what counts.
- Score it 0–10, on camera. Whether it ran, how close it hit the brief, how good it looks. I post the scoreboard inside the comparison guides on agentos.guide.
That's the whole loop. Every cell on Goldie Bench is one of those one-shot builds.
Why I built Goldie Bench
Because the AI leaderboard everyone reads — MMLU, HumanEval, SWE-bench — measures the wrong thing for people like us.
Those benchmarks tell a model vendor: "here's how to make the number go up." They're optimised against, marketed against, and they have nothing to do with whether the model can ship a thing for you in one prompt.
The buyer audience for these models — me, my community, you — needs a different question answered: can you ask the model for a playable game, a working simulation, a deployable page, in one prompt, and have it ship? That's the test that decides whether you'd actually wire the model into your stack.
Goldie Bench is that test. Every model gets the same fixed prompt. Every result is live and playable on this page. Every score is my own 0–10, posted publicly, with my reasoning attached. No vendor pays me. No score gets buried. If a model face-plants on one of my prompts, that face-plant is on the matrix forever.
What I'd love you to do next
If you find this useful:
- Read a comparison guide. Try GLM-5.2 vs Kimi K2.7 vs Opus 4.8 or The Three Dragons™. These are the source guides the scores on this bench come from.
- Watch a build. The YouTube channel has live builds nearly daily — most of these matrix cells were filmed.
- Join the community. AI Profit Boardroom is where 3,600+ founders are running the Agent OS daily. Templates, prompts, daily rooms, weekly walkthroughs. $59/mo, monthly.
- Get the Agent OS. The dashboard that runs every model on this bench under one roof. Inside AIPB.
— Julian
Run this stack yourself.
Every demo on this bench was built inside the Agent Operating System — one prompt, one shot, single HTML file out. The Agent OS, the prompts, the templates, the weekly walkthroughs and 3,600+ founders shipping with it every day all live inside the AI Profit Boardroom.