Скачать или смотреть Long Running AI Agents

This story is about a team who ran an AI coding agent, using Claude code, in a while loop so it kept working on tasks nonstop. They did this at a YC Agents Hackathon and wanted to see what would happen if they let an agent keep running forever. They set it up so the agent could read files, make changes, and push commits every time a file was edited.

The big idea was that if you give the AI a clear, simple prompt and let it run in a loop, it can port code from one language to another, like Python to TypeScript, and even keep an eye on GitHub issues for feedback. Along the way, the agent sometimes wrote extra tests, made improvements, and, interestingly, even stopped itself when it was finished or stuck. A few things they learned: short and clear prompts work best, the agent does not need a ton of instructions, and it helps to give it both the source and target code repositories plus a scratchpad directory for notes.

On the downside, running these agents for long periods can be expensive, less than $800 in one weekend's test, which is still cheaper than paying humans for the same work. They also noticed that a simple process outperformed more complicated setups, showing that sometimes “dumb and simple” just works.