The biggest thing is getting everybody in sync with the same models. Different agents are looking in different places for rules. Different agents respond better to different formats for rules. Unless you want like 4 separate repeated rules files, it really helps if everyone uses the same thing. The other benefit is the same models tend to make the same mistakes, so it's easier to catch issues as you learn the model. New models might be more powerful but if they respond very bad to the 20 hours you put into guard rails, it might not be saving you any time to switch. 

The other big thing is actually helping mentor people in how to prompt the agents. The agents are trained on human conversation. They respond well to conversational language and lots of context. They respond poorly when you are vague and imply you want them to engineer. One of the biggest things that leads to better code is being verbose and specific. 

If you want them to engineer or you need to be vague, put it in planning mode. If it's a big problem, have them write a todo file first. Agents are good at small specific tasks. Make sure by the time you're having them code that the instructions are specific. They lose context if they need to do too much at once. Walk through the planning with them, update the docs and plan in the file as you go along, then ask it to do one step at a time. You're basically writing out a feature and breaking it down into user stories. Then when you ask it to do a step in the file, it's going to create small subtasks rather than cram user stories into the actual agent tasks. You can clean up the todo files after. 

Then you need to make sure your team is doing the same thing. Don't be afraid to meet with the team (even if it's async in a group chat because people hate meetings). They might have other experiences with what works best for the agents you are using. The most important thing is that you're all in sync so you know what to look for in PRs and while working. Getting your team to be in sync on this is a soft skill you need to develop. Even without AI, getting your team to be consistent is important. 

People respond really well to positive reinforcement and praise. It's why all the AI models do it. When people are taking the extra time, make sure to call it out and thank them for it. The difference is you can make sure it feels genuine rather than the AI praising every little thing with hollow messages. 

At the end of the day though, you'll still have people that put minimum required effort in at work, and that's fine. It's Just a job and not everyone is gung ho about spending all their spoons on their day job. This is why minimizing the effort needed is important. Things like making sure the rules have instructions to update the rules and reminding people that it's only one extra prompt to get the AI to do it if it forgot. 

Tldr Everyone using the same things is often more impactful than everyone using their own wildly different best. Make sure agents are updating their own rules and docs. Try to make sure people feel good when they do it the team way.