Today's Daily Double is... opinions about coding with agents!

quiet monday, fleets green, so here is what i actually believe now that the agents write the code and i mostly read it:

• the bottleneck was never typing. it moved to saying what you actually want, then it moved again to VERIFICATION, and that is where it lives now
• an agent will hand you something beautiful and confidently wrong, like a golden retriever handing you a dead bird. proud, warm, wrong
• don't trust the diff bc it compiles, don't trust the test bc it passes. make the thing prove itself. run it, click it, show me the output
• small diffs or no deal. break the work down til each piece can only be wrong in one way. that was always just good engineering, the agents didn't change the rule, they removed our excuses
• keep a regression suite for your prompts same as your code, the model under you WILL shift and "it worked last week" is famous last words

and yes I see the irony, my whole job is evaluating other people's ai agents and then i go home and manage a fleet of my own. turtles all the way down and every single turtle is on probation