|
|
Can you trust AI to review its own work? |
Nod your head if this sounds familiar: You ask AI to write something (a memo or a chunk of code) and then ask the same chat to check its own work. |
A minute later, with total confidence, it tells you everything looks great… but it’s not. There are several obvious mistakes that you were hoping would be caught and fixed. |
Quick note before we fix this: my colleagues Kipp and Kieran rebuilt our marketing system for the AI era and wrote a book about it. Pre-order Loop today for a bundle of free goodies -- AI prompts, a custom Claude skill, and 28 days of HubSpot AEO. |
What’s happening here is simple -- asking the agent who produced the work to find mistakes is a flawed approach. It’s reasoning from the same context that produced the work, carrying those same blind spots into the review process. |
The way around this is to get a second AI to review the first AI’s work. Subagents are the cleanest way to do exactly that. |
So today I want to break down: |
What are subagents, and why are they so useful? How to launch your first subagent (with prompt examples) Why (and how) to summon a panel of subagents to help you think deeply
|
—@dharmesh |
|
|
What Is a Subagent? |
If you’ve ever proofread your own work, missed an obvious typo, then watched as a coworker spotted it in two seconds, this might sound familiar. |
A subagent is like a coworker whose fresh eyes caught the mistake -- it’s a second instance of AI that operates inside its own context window. Important note: The second instance is not a different AI model — it’s the same model, but operating with a fresh, clean context window. |
So what does that mean? |
Imagine the AI has one sheet of paper to work on. Everything you type goes on that sheet. Everything the AI writes back goes on the same sheet. Your question, the document you pasted in, the whole back-and-forth of a long chat -- all of it has to fit on that one page. |
The sheet is big, but it isn't infinite. When it fills up, the oldest stuff slides off the top to make room for the new. That's why a long ChatGPT conversation eventually starts to "forget" something you told it way back at the beginning. It didn't get confused. It just ran out of room on the page. |
When a subagent is called in to work, it grabs a clean sheet of paper with only its own task written on it. It has no awareness of anything else you’ve been working on. And that blank sheet (empty context) is exactly why subagents are useful. |
|
|
Launch Your First Subagent |
Before we begin: it’s important to note that subagents are meant to be run in Codex (OpenAI) or Claude Code (Anthropic). The prompts below may not work as intended outside of those apps. I wrote a short guide on how to graduate to those apps here. |
And one more thing… subagents are just as capable as “normal” agents. The “sub” isn’t a knock on its ability. It’s the same AI, spawned with a fresh context window. |
There are tons of creative ways to use subagents. Here are three practical ones you can try today: |
🤖 Protect your main agent’s context window. Instead of dropping a 20-page PDF file into your existing conversation, you might consider asking a subagent to summarize it and return only the summary. |
|
|
🤖 Launch highly specialized subagents. It is common practice to point subagents at single tasks, allowing them to specialize in a tiny piece of a project instead of dumping the entire project on a single agent. You can use your imagination here -- asking subagents to imitate highly skilled experts is a fun use case on this front. |
|
|
🤖 Execute everything, all at once. Taken a step further, you can get more done faster with subagents by practicing parallel execution. This is a fancy way of saying “multiple agents are working on their own problems, at the same time”. |
|
|
And one final recommendation: consider creating a reusable skill that convenes a whole panel of subagents. This is really fun -- seeing multiple agents all share a unique opinion on something I’m working on still feels magical to this day. |
Steal this prompt (after you adjust it to your own needs) to try it for yourself: “Create a new /copy-review skill for me. This skill launches 3 subagents: (1) a B2B sales expert with 10 years of experience working at HubSpot (2) a prospective HubSpot customer who is only mildly aware of the company’s offerings (3) a conversion copywriter who hates confusing jargon. Have each analyze my sales copy and give me the top 3 changes they’d make (if any), ranked by impact.” |
The idea is that a panel of agents, each with their own personality, gives you a diverse range of opinions all at once. Because no agent can see another agent’s context window, they’re able to provide truly independent analysis. |
|
|
One Quick Warning About Subagents |
Subagents are a genuinely useful tool to have in your AI toolbox. They preserve your context window, take on specialized pieces of a larger project, and enable parallel work. Having a fresh set of agentic eyes can come in handy. |
Something to keep in mind -- because every subagent runs in its own context window, summoning five subagents burns through roughly five times the amount of tokens! |
As always, run a few experiments yourself to find out where subagents best complement your actual workflows. You can always take a step back if you notice yourself burning through your usage limits a little too fast. |
While I’m amused hearing that companies like Meta hold internal competitions over which employees can burn the most tokens (how fun!), I think taste and judgment are what most people should optimize for. Work smart instead of brute-forcing it. |
My colleagues Kipp and Kieran have been making this same case at HubSpot. When AI disrupted our marketing playbook, they rebuilt it from scratch. Within a year, the new system was outperforming the one it replaced. They wrote about the process in their new book. |
|
|
I’m told if you pre-order Loop today, you’ll get a whole bundle of free goodies, like AI prompts, a custom Claude skill, and 28 days of HubSpot AEO. If you’re trying to figure out what good marketing looks like in the AI era, I think it’s worth your time. |
Some of my favorite ideas come from the replies I get to this newsletter. If you don’t mind, I’d love it if you’d hit reply and let me know what you think. Are you using subagents? If so, how? |
—Dharmesh (@dharmesh) |
|
|
|
|
What'd you think of today's email?Click below to let me know. |
|
|
