Use a computer agent to beat a web game
Ethan Mollick tasked Claude’s computer-control agent with finishing Universal Paperclips, and it mixed manual clicks, screenshots, and code before crashes interrupted the run.
- What they did
- Mollick opened a remote desktop session, loaded the game, and told Claude’s computer-use model to finish it. The agent watched the screen, clicked through turns, checked its own progress after roughly every 15 clicks, and experimented with pricing changes. After a crash, Mollick relaunched it with a hint to treat the desktop like a machine it could operate; the agent tried writing automation code, but when that failed it switched back to using the pointer and keyboard.
- How it went
- It carried out more than 100 actions on its own and stayed active for close to an hour without fresh input. A pricing calculation mistake ate up time and took several reminders to correct, and a second desktop crash ended the attempt with the agent saying it had won.
- Tips
- Expect this kind of task to move slowly and to burn time if the agent makes a small arithmetic mistake. It may also ignore attempts to steer it in the middle of a run, so set clear limits and require confirmation before anything irreversible.
Try it yourself with Claude computer use
Launch [browser game or web app] in the browser on the remote desktop and try to complete it. Begin by interacting directly with the page, but if you run into a repetitive task, see whether the site offers a safe way to automate it. Give me occasional progress updates, and before you do anything that cannot be easily undone, such as [payments/sending/deleting/submitting], pause and ask me to approve it. Keep your actions inside the game window, and stop when you finish or when you cannot make progress anymore; then explain what you tried.
Fill in the [brackets]. Agents can make mistakes, so review anything before it is sent, paid or deleted.
Get 3 agent uses like this every morning
With the day's top AI news. Free, and one click to unsubscribe.
Related agent uses
Build a browser-based 3D plane game
A Hugging Face product head asked Manus to make a Three.js plane game that runs in a browser.
Use spawned agents to build a Rust game engine
They had several Grok bots split up the work and produce a Rust editor with scripting, threading, and scene effects.
Use Claude to plan, code, and test a game tool
They used Claude Cowork over several months to design a game tool, generate code, and keep refining it through PowerShell, Maya, and screenshots.
Log NES state each frame for an agent
A developer is making an NES test rig so an agent can read emulator state and play games.
Run a Micropolis city with an agent
Hook an agent to a headless Micropolis simulator and let it manage a city from a simple goal sheet.
Build a bot for a real-time strategy game
They used LLM-written code to control bots in strategy-game benchmarks and had them compete in tournaments and ladders.