{"slug": "stop-making-your-ai-do-everything-itself", "title": "Stop making your AI do everything itself", "summary": "In a newsletter column, Matt advises AI users to split large jobs across subagents — helper AIs the main model spawns to handle individual pieces — rather than having one model do everything itself. Matt says subagents make work faster and often better, since each helper starts fresh with only its piece and the main AI reads only the summaries, and that cost can go down if simple tasks go to cheaper models or up if every subagent runs on the top model, as he does. He recommends the approach for research, document review and checking, but not for single cohesive outputs like one piece of writing or one app, and notes Fable 5.2 as an exception that works better as a single agent.", "body_md": "My advice on AI models has always been simple: use the best one you can get.\n\nBut even the best model has limits when it takes on a big job alone. It works through the job one piece at a time, and the more it has to hold in its head, the worse it gets.\n\nThe fix is something called a subagent. Get comfortable with them, and you'll get a lot more out of the AI you already use.\n\n## What a subagent is\n\nA subagent is a helper AI that your main AI creates to handle one piece of a job. It does that piece, reports back, and goes away.\n\nThink of a partner at a law firm. She doesn't read every document in a case herself. She splits them among associates. Each one reads a stack and brings back a summary, and she spends her own time on the judgment calls.\n\nSay you ask your AI to research your ten biggest competitors. Without subagents, it goes through them one at a time. With subagents, it creates ten helpers, gives each one a competitor, and they all work at the same time. Then the main AI reads the ten reports and writes yours.\n\n## Why it's worth it\n\n**It's faster.** Ten helpers working at once finish long before one AI working down a list.\n\n**The work is often better.** An AI can only keep so much in its head at once. Make one AI read forty documents, and by the end it's skimming. Each subagent starts fresh with just its piece, and the main AI only has to read the summaries.\n\n**It can cost less, or more.** The main AI can hand the simple pieces, like searching and reading, to a faster, cheaper model, and keep the planning and the final call for itself. That saves money. But plenty of people, me included, run every subagent on the same top model, which can cost more. We do it because the result is better.\n\n## When not to use them\n\nSubagents are for work that splits into separate pieces: research, reading through piles of documents, checking things.\n\nThey're the wrong tool for one thing that has to hang together, like a single piece of writing or one app.\n\nIt also depends on the model. Most models are really, really good with subagents. Fable 5.2 is the exception: it just works better as one agent. In my Fable review, I mentioned that when I had several copies of it build a game as a team, the results were mediocre. When I let a single one do the whole thing, it took off.\n\nSmall jobs don't need them either. A team of helpers usually uses more of your plan than one AI working alone, so save them for the big stuff.\n\n## How to start\n\nYou need a real agent for this, like Claude Cowork, Claude Code or Codex.\n\nThen just ask. On your next big job, add this to your prompt:\n\nThe middle of that prompt matters most. A subagent starts from zero. It only knows what the main AI tells it. It's the intern test again, except this time your AI is the one writing the note.\n\nTry it on the next big job you hand your AI. Hit reply and tell me what you gave it.\n\nBy the way, if you write software, [Devin Bots](https://somethingbig.ai/go/devin-1117?p=web-text) do this for you. Give one a goal and it splits the work across a team of Devins on its own.\n\n— Matt\n\nWhat should I write about next? [Vote or suggest a topic →](https://somethingbig.ai/next)", "url": "https://wpnews.pro/news/stop-making-your-ai-do-everything-itself", "canonical_source": "https://somethingbig.ai/stop-making-your-ai-do-everything-itself", "published_at": "2026-11-17 17:00:00+00:00", "updated_at": "2026-11-17 20:16:23.284658+00:00", "lang": "en", "topics": ["ai-agents", "artificial-intelligence", "ai-tools"], "entities": ["Matt", "Claude Cowork", "Claude Code", "Codex", "Fable 5.2", "Devin Bots", "Devin"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/stop-making-your-ai-do-everything-itself", "markdown": "https://wpnews.pro/news/stop-making-your-ai-do-everything-itself.md", "text": "https://wpnews.pro/news/stop-making-your-ai-do-everything-itself.txt", "jsonld": "https://wpnews.pro/news/stop-making-your-ai-do-everything-itself.jsonld"}}