My honest take a year into building AI copilots for top YC startups

I’ve spent the past year building AI copilots for seed to 500-people companies, 5+ of which are YC startups.

6 months ago we were seeing autonomous agents, v0/lovable style chats, and product knowledge agents brought into production. Almost everyone is now pivoting into AI-native applications, and 90% of the top angels’ AI investments target the application layer. Here’s 4 reasons why:

1*. The more valuable the work, the more you need human in the loop*

I know you love the sci-fi vision of AI agents doing entire workflows for us, tbh so do I (it’s coming)

But here’s the truth: If you’re automating work, it should be work that’s important enough to be worth reviewing.

If someone is willing to let AI do the work completely unsupervised, it’s probably not very valuable to them. You might let an agent look up plane tickets, but would you give it access to your wallet to buy them without reviewing? Probably not.

https://imgur.com/a/DdBqc8q (code is open-source)

I do think this will change as AI gets better, but frankly agent’s just aren’t ready yet

2. UI > Text.

Look, I’m a lazy guy. I see paragraphs of text and my eyes just glaze over. The average attention span has dramatically shortened, and paragraphs of text just aren’t cutting it.

If you’re going to do human in the loop, leverage your UI.

Don’t make your AI give big paragraphs of text. Show the user what the agent is doing! Directly make changes in your app that the user is already familiar with.

3. Working solutions are 90% software and 10% LLM.

Ironically what we’re seeing is that pure LLM solutions don’t have that much of a moat. You can spend hundreds of hours fine-tuning your model, or create superior agent workflows to your competitors, and it gets leapfrogged by the next model release.

Software is still more consistent, cheaper, and has superior infrastructure (at least for now). Instead of thinking “What’s the craziest agent workflow”, think “what is something that is almost possible, but AI fits that last puzzle piece?”

4. Normal people don’t understand how to use AI. Applications give you context.

Using LLM’s is hard. It takes good prompting structure, copy and pasting important context, and knowledge of what to ask the agent.

In an application, you already have the most important context. You already know what the user is trying to do, and can automatically pull whatever data you need if you need to.

Think of Cursor. When you ask for something, it can automatically search through files and code to do what it needs.

I’m sure you know all the options for building the agent itself – Mastra, Langchain, Simstudio, etc. etc.

The frontend space is less well established, but if you’re looking for just a chat w/ custom message rendering, you can use something like AI SDK or assistant-ui. If you’re looking for something deeper that helps with agent reading & writing to state, context management & voice, I use Cedar-OS (it is only for react though) for customer work.