{"slug": "use-your-subagents-as-judges", "title": "Use your subagents as judges", "summary": "A newsletter post from Matt at SomethingBig.ai recommends using fresh, blind subagents as judges to critique AI-generated work, arguing that AI grading its own output gives itself inflated marks the way a student would. The post advises giving each judge a specific persona, instructing it to be harsh, keeping it blind to who produced the work, and demanding concrete feedback rather than a score alone, with applications to writing, games, apps, plans and research. It says the technique requires a real agent such as Claude Cowork, Claude Code or Codex.", "body_md": "Remember how I told you about subagents on Tuesday? If not, [click here](https://somethingbig.ai/stop-making-your-ai-do-everything-itself).\n\nMost people think of subagents as a way to split up a big job. But there's a more advanced way to use them: as judges, to check work.\n\nWhy bother? Because AI does its best work when it gets feedback and takes another pass. It's like a student and a teacher: the student hands in a draft, the teacher marks what's weak, and the next version is better.\n\nBut the feedback has to come from someone else. Give a kid an assignment and ask them to grade it themselves, and they'll probably give themselves an A. Have someone else grade it, and they might not.\n\nAI is the same. Ask your AI to grade its own work and it will give it a better grade than it deserves. It made the thing, so it reads it the way it meant it, not the way it really is. A fresh blind subagent hasn't seen any of that. It didn't do the work, it doesn't know who did, and it has no reason to go easy.\n\nThe prompt I gave you on Tuesday ended with a small version of this: \"Then have a fresh subagent check the final result against my original request.\"\n\nHere's the bigger version. Once your AI has finished a task, any task, say:\n\n\"Ok, now launch a fresh blind subagent to critique this as a real person would, and give feedback on how to make it better.\"\n\n## What you can judge\n\nPretty much anything your AI makes.\n\n**Writing.** Have a few subagents read your email or report as different people, and tell you the exact line where they got bored and stopped.\n\n**A game or an app.** Have a fresh subagent actually play the game, or click through the app, as a first-time user. Where did it get stuck? What was confusing? What was boring?\n\n**A plan.** Have one read your business plan as a skeptical investor and find the weakest part.\n\n**Research.** Have one check the report against the sources and flag anything it can't back up.\n\n## Four things that make the judges better\n\n**1. Give each judge a specific person to be.** A busy CEO, a skeptic, a kid playing your game for the first time.\n\n**2. Tell it to be harsh.** Otherwise it'll be polite, and that doesn't help you.\n\n**3. Keep it blind.** Don't tell it who made the thing or share your conversation. If it knows it's yours, it goes easy on you.\n\n**4. Don't settle for just a score.** A score is fine, but on its own it doesn't tell you what to fix. Have it give real feedback too, like \"I got stuck on the second screen because the button was hidden.\"\n\n## The prompt\n\nSame as Tuesday: you need a real agent for this, like Claude Cowork, Claude Code or Codex.\n\nOnce your AI has finished something, paste this in:\n\nSwap in the people who matter for what you made.\n\nTry it on the next thing your AI makes. Then hit reply and tell me what your judges said.\n\n— Matt\n\nWhat should I write about next? [Vote or suggest a topic →](https://somethingbig.ai/next)", "url": "https://wpnews.pro/news/use-your-subagents-as-judges", "canonical_source": "https://somethingbig.ai/use-your-subagents-as-judges", "published_at": "2026-11-18 01:47:08+00:00", "updated_at": "2026-11-18 02:11:07.472101+00:00", "lang": "en", "topics": ["ai-agents", "ai-tools", "large-language-models"], "entities": ["SomethingBig.ai", "Matt", "Claude Cowork", "Claude Code", "Codex"], "also_reported_by": [], "takeaway": "Have a fresh blind subagent critique finished work in a named persona, told to be harsh and to give line-level feedback rather than a score, because self-grading inflates results.", "recommendation": {"advice": "After your agent finishes a task, launch a fresh blind subagent with a specific persona, instruct it to be harsh, and require concrete feedback on what to fix rather than a score alone.", "playbook": "agent-architecture", "playbook_url": "https://wpnews.pro/playbooks/agent-architecture"}, "alternates": {"html": "https://wpnews.pro/news/use-your-subagents-as-judges", "markdown": "https://wpnews.pro/news/use-your-subagents-as-judges.md", "text": "https://wpnews.pro/news/use-your-subagents-as-judges.txt", "jsonld": "https://wpnews.pro/news/use-your-subagents-as-judges.jsonld"}}