Posts

Do not ask whether the agent finished the code. Ask how it proved the feature works.

We've observed this multiple times in various projects we've built.

While building a web app, we had to use browser testing to verify the user interface interactions created the correct database fields.

While building a video game, we set up a GPU-enabled cloud workstation so the agent could play the game with new features to prove they worked.

While building an SMS-based AI assistant, we gave our coding agent access to an SMS API so it could send messages to staging builds of our assistant agent. Agents talking to agents!

Don't trust your coding agent if it just wrote the code. Make it show proof by using the application.

Building that verification environment can require as much thought as building the feature itself but it's worth it.