The author had GPT-5 in ChatGPT agent mode play online versions of Math Games with Bad Drawings, and it mostly worked, unlike earlier rental car and tax form attempts.
The author had GPT-5 in ChatGPT agent mode play online versions of Math Games with Bad Drawings, and it mostly worked, unlike earlier rental car and tax form attempts.
The author pointed GPT-5 in ChatGPT agent mode at online versions of games from Ben Orlin's Math Games with Bad Drawings. The agent clicked around the game sites in its own browser and chose moves, with its reasoning visible. The author played the opposing side and had to type "your turn" after each move, since the agent couldn't tell when it was its turn.
Basically worked, which beat earlier attempts at rental car booking and tax forms. It played poorly, though. The author beat it easily as both Bob and Alice, and it wrongly claimed it could watch for turn changes.
The agent can't watch a page for changes, so you have to prompt it every turn. A commenter also noted the games may be in its training data.
Open [URL of an online game, e.g. a math or puzzle game] and play it for [number] rounds, trying to get the highest score you can. Tell me your score and describe where you struggled. Stop if the site asks for login, payment or personal information.
Claude Code got an old DirectX game running on modern hardware, decompiled much of it, built a headless parallel runner, and now writes full level guides with screenshots.
ChatGPT Agent Mode opens a real browser and books appointments by clicking through the booking site; the post gives a setup guide.
A developer used ChatGPT Agent mode to extract revenue growth, free cash flow and net cash from 10-Q filings with source URLs, feeding a multi-agent valuation and sentiment workflow.