When coding with agents, if you don’t add a feedback loop, you’re the feedback loop. You become QA, and telling the agent that something still doesn’t work is extremely frustrating.
So take yourself out of the loop. After all, the way we (used to?) program is to “write a little, test a little.” Give your agent the power to iterate and correct its own mistakes before it comes back to you.
For example, have the agent:
- Write an API endpoint, test it with
curl, and inspect the database’s new state - Build an invokable script or CLI for a small piece of functionality and use it to test and iterate
- Use
agent-browseror Codex’s Computer Use to literally test out the app by pointing and clicking
Agents are good in a single blind one-shot generation, but not that good. They’ll make mistakes — just like you would. The only difference is that you can test your work with frequent feedback loops. Your agent should be able to do that too, if only you provide the right direction.