the problem kept shipping agent output that looked right and wasn't. no way to tell the difference until a human read every line. the fix anthropic's outcomes feature, announced at code with claude sf in may 2026, adds a separate grading agent. you write a rubric describing what "good" means for the task, and a second instance of claude scores each run against it before anything gets delivered. action write the rubric before you build the agent, not after. list 3-5 concrete pass/fail criteria for the output, then let the grader enforce them on every run. tradeoff pro: turns "does this feel right" into a score you can track and improve against. con: a vague rubric gives you a confident, meaningless score, the grader is only as good as the criteria you wrote. takeaway stopped eyeballing agent output. wrote five criteria, the grader caught what i was skimming past.

people keep asking what's actually in my ai stack, so here's the honest breakdown. claude's about half my usage, writing, editing, thinking things through out loud. chatgpt's maybe a third, mostly for doing rather than thinking, quick research and image stuff. perplexity when i actually need a citation instead of an opinion. gemini only when i'm already stuck inside google docs or gmail, otherwise i forget it exists. but the split isn't the part i'd actually recommend copying. it's what i don't do: no single tool for everything, no first draft goes out unedited, nothing gets published without reading it out loud first. the percentages will look different in six months. that last part probably won't.