When you stop and wait for an LLM it breaks flow.

If you stop and review what an agent did, it breaks flow.

Personally, my experience has been the best with doing this:

- Start coding

- On a second laptop, run an agent in yolo and tell it to periodically pull changes and if there are any `AGENT-TODO:` items in the code, do them.

- As you code, if you find something irritating or boring, `AGENT-TODO: ...` it.

- Periodically push your changes.

- Periodically pull any changes your agent has pushed down.

- Keep working; don't stop and check on the agent. Don't confirm every action. Just yolo.

If that's too scary, have it put up PRs instead of pushing to the live branch. /shrug

...but, tldr: if you're sitting there watching an agent do things, you're not in flow. If you're kicking multiple agents off and sitting waiting for them, you're more productive, but that is absolutely not flow state.

Anyone who thinks they are, doesn't know what flow state is.

The key to maintaining flow is having a second clone, or a second machine or something where you can keep doing work after you kick the agent off to do something.

(yeah, you don't need a second laptop, but it's nice; agents will often run things to check they work or steal ports or run tests that can screw with you if you're on the same machine)