I watched an AI hire its own team last night

I gave an AI a goal last night, and a few minutes later, it had hired a lead engineer.

It wrote a job description, picked a tech stack, defined the role, brought in a lead engineer agent, and started building.

Then it decided it needed a product manager, so it created the role and hired the agent. Then a marketing person. Same thing.

Within minutes, I was watching a small company form itself around a single goal I'd typed in.

The project is called Paperclip. You give it a business goal, it creates a CEO agent, and that CEO starts making decisions: what to build, who to hire, what stack to use.

The "employees" are all agents that actually do the work. I'm still early with this, still testing it — but sitting there watching it happen was one of those moments where you go: oh. This is real now.

I know what you're thinking — cool demo, but is it practical? Honestly, not fully, not yet. But that's not really the point.

The point is that the building blocks are here. Claude Code is getting capable enough that these multi-agent systems actually work. Not perfectly, but they work. And this connects to something I keep coming back to in this newsletter: the future of working with AI isn't one chatbot doing one thing. It's ecosystems.

Think about it like this. Claude Code today is a marketplace — it has plugins, and inside those plugins you have skills, connectors to different tools, and agents that handle specific tasks, all of which can talk to each other.

If you're building anything with AI, this is the framework: marketplaces, plugins, skills, agents — for your business, your team, your workflows. Paperclip takes that a notch further, with multiple agents working together toward a shared goal and an AI org chart that builds itself.

I've been bouncing between email tools for months. Superhuman has an incredible UI — the mobile app is genuinely the best I've seen — but their AI features have been lagging behind.

I tried Jace for a while, and the AI was better, but the interface and overall experience weren't quite there.

Then last night I found EXO. It's open source, built by Y Combinator partners, and here's the wild part: it's basically Claude Code living inside your inbox. You bring your own API key, spend your own tokens, and Claude sits there reading, drafting, and organizing.

I tested OpenCode before, which is essentially an open source version of Claude Code. At first I thought, why bother when Claude Code is this good? But I'm rethinking that. Open source models running locally are getting genuinely impressive — not Claude-level yet, but good enough for a wide range of tasks, and improving fast.

Here's where my head is going: in a few months, I think the smart move is to get some serious hardware — GPUs, maybe a Mac Studio or two — run a strong open source model locally, and plug it into tools like OpenCode, EXO, and Paperclip. Your own AI infrastructure, your own tokens, your own agents running 24/7 on hardware you own.

That's not science fiction. The pieces exist today; they just need a little more time to mature.

It's April 1st, Q1 is done, and I'm sitting here thinking about how fast all of this is moving. If you're reading this and you still haven't tried Claude Code, or Notion Agents, or even just switching models on ChatGPT (please stop leaving it on "auto") — you're missing out. I say that with love, but it's true.

Every newsletter I share three to five concepts, and you don't have to try all of them. Just one. And if you're unsure where to start, reply to this email — I read every reply and I'll point you in the right direction, no pitch, just help.

Every single person I've shown Claude Code to has gotten hooked within the first session. The ones who pushed past the initial discomfort are the ones having fun right now, while everyone else is still deciding whether to try.

Q2 is a good time to stop deciding.

I'm stateside for the next couple weeks.

San Francisco April 5th-12th, then Atlanta until the 19th. If you're in either city and want to grab coffee and talk AI, hit reply.

Always happy to meet readers in person!

Thanks for reading,

Tim