How I run my businesses and my life on a fleet of AI agents

7:14 on a Tuesday morning, kitchen, kids eating breakfast. A billing question comes in about one of my companies and I don’t want to lose it. I open Telegram, hold the mic button in the Tabs topic, and talk for fifteen seconds: “the May invoice for the Henderson account looks like it double-billed the platform fee, can you pull the line items and draft a reply to their AP person.” I put the phone down and pour the milk. By the time I’ve sat down the agent has read the relevant thread, pulled the invoice, found the duplicate, and dropped a draft email in front of me with a button to send. No laptop. No terminal. I never opened a single app that looks like software.

That moment is the whole point of this post. The thing answering me is Claude Code, Anthropic’s coding tool. I am not using it to write code. I am using it to run my morning. And the reason that works at all is a piece of software I built to fix the one thing wrong with Claude Code, which is the way you have to talk to it.

The thing nobody tells you about Claude Code

I have tried most of the agents. Cursor, Aider, Gemini’s CLI, OpenAI’s Codex, a handful of the open-source ones you can host yourself. I ran each of them against real work, not demos: shipping code, multi-step research, triaging my inbox, managing projects across three companies. The post next door to this one, Why I built Neutron, has the side-by-side if you want the receipts on each.

Here is the short version. Claude Code won, and it wasn’t close.

The surprise is in the name. It’s called Claude Code, and everyone reads that as “a tool for programmers.” It is the best tool for programmers I’ve used. But that undersells it badly, because it is also the best general-purpose agent I’ve used for everything that has nothing to do with code. The coding framing is a marketing accident. What Anthropic actually shipped is the deepest, most capable agent runtime on the market, and they happened to point it at developers because developers were the first people willing to pay for it.

What makes it deep isn’t the model alone. Every company has a good model now. It’s the environment around the model: the ability to use tools, to remember instructions, to hand work off to helper agents, to run unattended, to plug into your email and calendar and files. Claude Code has years of that built in and it’s miles ahead. The other tools have a smart model wrapped in a thin shell. Claude Code is a smart model living inside a real workshop.

So I made a bet most people building in this space would call backwards. I stopped looking for the perfect agent and decided Claude Code already was it. Then I went to work on the only thing actually holding it back.

What it looks like to run your life on a “coding tool”

Once you stop thinking of it as a code tool, the obvious question is what else it can do. For me, the answer turned into nearly everything I do at a desk.

I run three businesses through it. Each one is its own long-running conversation that never forgets the history: a running status file, a memory of every decision, the actual repository of files behind it. When I need to think through a pricing change or chase down why a number moved, I’m talking to something that already knows the context. I’m not re-explaining the business every time.

I run my side projects through it. This website. An art and music project I’m building gear for. A course I’m helping write. Each one gets its own space and its own memory.

I run my personal life through it. Calendar triage, reminders that actually know what’s going on, a running log of my health protocols, family logistics, the notes and ideas I’d otherwise lose. When I think out loud about something, it gets filed where I can find it later instead of evaporating in a chat thread.

And yes, it still writes code, which is the one thing it was advertised to do.

The point is not that I found clever hacks. The point is that the most capable general-purpose assistant available right now is sitting inside a product labeled “for coding,” and most people will never try it for anything else because the label told them not to. I tried it for everything. It’s the best at everything.

The one thing wrong with it

Here’s the catch, and it’s a real one.

Claude Code lives in a terminal. If you’re not a programmer, a terminal is the black window with the blinking cursor that hackers use in movies. It is beautiful for what it was built for: one person, sitting at one computer, working on one project, with their hands on a keyboard and their full attention on the screen.

That is not the shape of my life, and it’s probably not the shape of yours either.

• I am not at my laptop for most of the day. I’m in the kitchen, in the car, at a kid’s swim lesson, walking the dogs. The terminal can’t come with me.
• I want to talk to it, not type at it. Typing a paragraph into a terminal on a phone is a special kind of misery. I’d rather hold a button and say what I mean.
• I run a lot of these at once. Right now I have around two dozen separate Claude Code sessions going, one per project, each doing its own thing. A terminal shows you one at a time. I need all of them, from wherever I am.
• I want it to reach me, not just answer me. Email comes in, a calendar event changes, a reminder is due, a long task finishes while I’m asleep. The terminal can’t tap me on the shoulder.
• I want it to remember. Every fresh terminal session starts from zero. I don’t want to re-explain who my COO is, or how I like emails drafted, or what we decided last week, every single time.

None of these are flaws in the agent. The agent is excellent. They’re flaws in the doorway to the agent. The intelligence I wanted was already there. The way in was wrong.

So I built a better way in. I call it Neutron.

What Neutron actually is

Neutron is a layer that sits on top of Claude Code and gives it the interface I always wanted. The agent underneath is unchanged. I just replaced the terminal with something I can actually live in.

Out of the box that something is a web app and a mobile app: one screen with a list of projects, each one a live session you can talk to. That is the default, and it is all most people need. I happen to run mine through Telegram, the chat app, because it’s already on my phone, my laptop, and my watch and I’m in it all day anyway. Telegram is an optional way in, not the product. I’ll describe my Telegram setup below because it’s the one I actually live in, but read “Telegram” as “whatever window you point at the fleet.” Here’s what that buys me.

It’s just chat, so it’s mobile-first by default. The interface is already on my phone, my laptop, my watch. I don’t install anything heavy or learn anything new. I just send messages. Voice messages work the way they always do, so I can talk to my whole operation while I’m doing something else with my hands.

Each conversation is its own project, and each one remembers. The interface splits into separate threads, one per project (in Telegram these are forum topics; the web app has the same shape). I have one per business, one per side project, one for my personal life, and so on. Each thread is wired to its own dedicated Claude Code session running back on my Mac. Open the Tabs thread and you’re talking to the agent that knows everything about that business and nothing about the others. The projects never bleed into each other. The work stays clean.

It picks up exactly where it left off. I can fire off a voice note from my phone at a stoplight, then sit down at my desk an hour later and keep the same thread going on my laptop. The conversation didn’t reset. The state lives on my machine and the chat interface is just the window I happen to be looking through. Desktop and mobile are the same session, always in sync, because there’s nothing to sync. There’s one brain and many windows.

It remembers across time, not just within a conversation. This is the part that changed the most for me. Neutron adds a long-term memory layer the raw tool doesn’t have. The agent remembers my preferences, who the people in my life are, the history of each project, what we decided and why. I stopped re-explaining myself. That alone gave me back an enormous amount of friction I didn’t realize I was paying.

It handles its own housekeeping. Conversations that run for weeks would normally overflow and lose the thread. Neutron checkpoints them so they keep their memory without falling over. I never think about it. It just keeps working.

It can reach out, not just respond. I can tell it “remind me every morning at 10 about the dogs” or “ping me at 6pm with how Tabs did today,” and it does. The reminders aren’t dumb alarms. Each one is a tiny fresh agent that wakes up, reads the current state of things, and writes me a message that actually fits the moment, with the weather, my calendar, whatever’s relevant that day.

It can send work off to do in the background. When something’s going to take a while, like a deep research job or a code build, the main agent hands it to a helper that goes off, does the work, and comes back with the result posted to the right topic when it’s done. I don’t sit and wait. I get a notification later.

That’s it. No new app to learn. The smartest agent I’ve found, reachable from my pocket, in a chat thread, by voice, with a memory.

What a day actually feels like

The clearest way to explain it is to walk through a normal day.

The billing thing from the top of this post is the morning shape: a fifteen-second voice note in the right topic, and a finished draft waiting for one tap to send. The agent did the reading, the math, and the writing. I did the deciding. That’s the division of labor I want with every tool: it does the grinding, I do the judging.

A bit later a morning brief is already sitting in my general topic. Something composed it before I woke up by reading my overnight email, my calendar, the weather, and where each of my projects stands. It tells me what’s actually on me today and what’s at risk if I ignore it. The kind of summary that used to cost me forty-five minutes of clicking around now costs me three minutes of reading.

During the day, if there’s code I want built for one of my software projects, I type one line into that project’s topic and walk away. A build-and-review loop runs on its own and a notification shows up later telling me it’s done. I wrote about that pipeline in detail in a separate post. From where I’m standing it’s just: ask in chat, get a result, go do something else.

When I think out loud about a decision in any topic, the system quietly catches the framing and files it into a personal knowledge base I can search later. Half-formed thoughts I’d normally lose become something I can pull back up months from now. The memory post goes deep on how that works if you care about the machinery.

And at the end of the day I get a short digest of everything that ran while I wasn’t watching, before I go to bed.

None of those moments involve a terminal. None of them involve me at my desk. Most of them involve me talking, not typing. That’s the entire experience I was after.

The substrate and the interface

If there’s one idea to take from this, it’s the split between two things people usually lump together.

The substrate is Claude Code. That’s the engine, the intelligence, the thing that’s genuinely the best on the market right now and is going to keep getting better without me lifting a finger, because Anthropic ships the improvements. I bet everything on that engine and I’d make the same bet again.

The interface is Neutron. That’s the part that makes the engine usable for an actual human living an actual life: on a phone, by voice, across a couple dozen projects at once, with a memory that holds, with reminders that find you and work that runs while you sleep.

The combination is what makes Claude Code something you can genuinely live inside, not just the one project you happen to have a terminal open for. The engine was already great. It just needed a door you could walk through carrying your real life in your hands.

If any of this is interesting for your own work, Neutron is open source and self-hosted: the code is at github.com/rjunee/neutron (Apache 2.0), and neutronagent.ai is the hosted version if you’d rather not run a gateway yourself. Neutron is open source is the announce and the self-host path. Why I built Neutron is the longer origin story. And if you want help thinking through agents for your own business, that’s what my consulting is for.

I'm productizing this substrate as Neutron for operators who want a real agent system without rebuilding it themselves. Separately I take on a small number of consulting engagements per quarter for teams shipping into production. Services →

← All posts