Researchers had OpenAI's Computer-Using Agent play NYT Wordle; it solved only 5.36% of about 200 runs, mainly from misreading tile colors.
Researchers had OpenAI's Computer-Using Agent play NYT Wordle; it solved only 5.36% of about 200 runs, mainly from misreading tile colors.
The researchers pointed OpenAI's Computer-Using Agent at the New York Times Wordle game. The agent controlled it through a standard computer interface, the way a person would. They ran it over a week of daily Wordle puzzles, across several hundred runs. They watched its behavior to find where it failed.
The agent solved only 5.36% of runs. The main failure was that it recognized tile colors inconsistently, and its accuracy depended on the context.
Open [Wordle or another browser word game] and play today's puzzle. After each guess, read the tile colors carefully and state what each letter's color means before choosing the next guess. Stop when you solve it or run out of guesses, then tell me your guesses and any color you were unsure about.
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.
Four Grok 4.20 agents built a tower defense, mini Factorio, CHIP-8 emulator and roguelike shooter in parallel over a day.
Grok built the browser game Roofline, then trained a PPO agent that plays the live page, reaching a best score of 39,359.