The AI tools got a lot better in recent months and the change is enough that almost everyone should adopt a "swarm mentality" when thinking about tasks.

1. Like Interns

Imagine 20 interns from MIT just arrived at your office. You can't tell them to wait a week while you collect your thoughts, no - you give them every task you can and tell them to come back with questions. It's the same with coding agents - throw every task at them, all at once.

2. Imperfect Results

Not everything will immediately work or turn out as you expect, that’s ok and just like running a team of humans. Plus, you can only run a limited number of tests concurrently, so the attrition does your selection work.

3. Testing

I suggest 3 levels of testing, ask the coding agent to build its own unit tests. Next create a separate agent to analyze the repo looking for problems such as overfitting, extraneous code, bad tests etc. Then have a digital twin replicating the real world that is ringfenced from the other agents and acts like a QA team.

Best of luck out there.