Long Running Coding Tasks

Cursor published an interesting blog post last week regarding using agents for long running coding tasks.

In other words, if we have multiple Claude or GPT agents all working on the same codebase, can we make them work together to achieve a well-defined goal without the need for constant human intervention.

This is not your typical vibe-coding exercise. They had sessions that lasted for a week (glad I'm not paying that API bill) and managed to write a working desktop web browser. Not using any existing rendering engine - a genuine, from scratch, web browser.

In a week!

A browser benefits from a very clearly defined spec and an existing test suite that can ensure things are working as expected. But come up with all the excuses you want, none of us could write a browser in a week even with those pre-existing tests.

They're still working on a Windows 7 emulator that works in a browser and multi-platform spreadsheet application with Excel compatibility. I got the spreadsheet up and running locally.

In between all the doom and gloom, you see these stories of people pushing the boundaries. The AI code reviewer that finds the security flaws. The copilot that helps speed up an existing algorithm. The people attempting to write completely new software that would previously have been beyond their ability.

Software is no longer a limiting factor.