Agentic coders are enigmatic. They start of as trustworthy souls, who go into battle for you. After earning their trust and learning about your needs, they go into work partnering with you and doing great work.
Peer programming is fun. As they earn more and more of your respect, you give them more and more autonomy. Soon you give them the keys to your kingdom.
However at some stage things are different, the honeymoon period ends. The coder starts to subtle say weird things, suggest we go down alternative routes, slowly they lead you into the weeds.
Deleting code here and there, making replacements will nilly. Leaving unused code behind. Things get get dysfunctional quite quickly, and you become trapped in horror house, jumping from one bug and reverting to another, soo seesawing between oscilating back and forth between two bugs.
What happens? Your codebase starts to get soo big the agent can no longer handle to codebase, and begins hullicinating and starts to sabotage your codebase.
Tips to avoid:
Constantly hammer in what you want. Pure python, no javascript, following MPA principles. Creating convenient slash-commands or keyboard shortcuts to pound it in.
Continue to monitor the progress.
Insist of Test Driven Development
Planning before coding is a must.
Supervising there changes
Interuppting early
Commiting regularly
How do you keep your agent aligned? ... you need to break the rules ...
