StartBuilding
Run live before it reaches you.
Not a prompt that looked right on paper. 37 of the 40 agents were pasted into a live model exactly as you will paste them, and given three realistic business situations: the everyday case, the trap where the right answer is “I don’t know”, and the awkward edge. 111 scenarios in all, last run on 5 November 2026. The other 3 read the live web, which a test run cannot do, and are not in it.
37agents run live
111live scenarios
123real replies you can read
3tools it runs in
What that gives you
- A real answer under every check. Inside the Build tool, each check carries the reply the live run produced for exactly that input. You compare your screen with one that was right, instead of guessing.
- Prompts tightened on evidence. Where a scenario did not pass, the rule the model bent was made plainer and the scenario run again. The prompts you receive are the ones that came out of that, not the first draft.
- Tested on a small model. The runs use Claude Haiku, the smallest model in its family, so the prompts were tightened against a weaker reader than the free ChatGPT, Claude or Gemini you will use.
- Then built on film. 37 of the 40 agents were built end to end in a free ChatGPT account, all three checks run live, and recorded. That film sits on the agent's Build page.
- 3 agents are not in the run. They read the live web, which a test run cannot do. Their pages say so, and the three checks you run yourself with your own pages are the test.
What it does not promise
A model is not a calculator. A few scenarios come down to arithmetic, working out which date “next Thursday” is or counting characters, and small models get those wrong some of the time whatever the prompt says. The checks tell you where to look. Your own run of the three checks, in your own chat, is the pass that counts, which is why the Build tool asks you to mark it yourself.
Runs in
ChatGPTClaudeGemini
See the agents →