Also to add another point is that if you felt like an agent did not help you correctly, or way overshot, did too much edits, etc. Go back to the original prompt, rephrase it - sometimes you need 1-2 times. Sometimes the model just don't work for your workflow. It can become quite delicate.

One of the bigger things is when you introduced some bug, start working backwards with the agent, simplifying whatever you build to its bare necessities, and the moment it dissapears, start a new chat, and build it back up to what it was before (in the desired non-bugged state). This often works if you then also switch to a completely different model.