How I work
The machine, the harnesses, the gated automations, and the one decision I make thirty times a day. Everything else is scaffolding.
My job used to be doing the work.
Now it's deciding which bucket the work goes in.
That sounds like a promotion. It isn't. It's a different job with a different failure mode, and it took me longer than it should have to notice the switch had happened.
People keep asking how I get through it all - Automaze, the books, the podcast, the side projects, the stuff I build that never ships. Fair question. So here's the honest answer, including the parts that aren't replicable.
The whole system, in one line
I make 20 or 30 decisions a day and almost all of them are classifications.
Rails or agents. Bespoke or general. Human, automated, somewhere in between - or time to kill the process entirely.
That's it. That's the job.
Everything else you're about to read - the machine in my office, the harnesses, the briefings, the VIP list - exists to protect the context in my head so I can keep making those calls all day without the quality degrading. Bezos said he gets paid to make a small number of high-quality decisions. I'm not running Amazon. But that's the target, and everything else is scaffolding.
Automated doesn't mean AI
Here's the reframe most people need.
Anything I've done more than a few times, and have a process for, gets automated. But automation sorts into three buckets, and only one of them is agentic:
Deterministic. Known inputs, known outputs. No model anywhere near it. Just automation.
Hybrid. A fixed pipeline with one fuzzy step dropped in exactly where judgment is needed. Write the actual email. Make the call at this branch. Everything before and after stays dumb.
Agentic. Genuinely fuzzy end to end. An agent owns the whole task.
Most of what I run is the middle one.
The instinct to make the entire workflow intelligent is usually wrong. It's slower, it costs more, and it's much harder to debug. A deterministic pipeline with one smart step is easier to reason about than an agent that's smart the whole way through and unpredictable in fourteen places.
Boring where it can be boring. Smart where it has to be.
Why almost everything is gated
Most of my automations require human approval at some stage.
That's not caution about the technology. I'm all in on the technology - I spend tens of thousands of dollars a month on tokens. It's accuracy about the domain.
Every client we serve is an edge case. The processes look nearly identical between them and then diverge in exactly the place that matters.
So the dial between fully automated and human-gated is technically a setting, but in practice it's a judgment about how we want to work. AI takes the dirty work off my plate. It doesn't stand in for me in client communication.
The setup
Local or cloud is a real debate and both camps have a point. My answer is that my local environment is the cloud.
There's a dedicated machine in my home office, accessible to me and only me, from anywhere. Fully configured the way I want it. Safe to run experiments on. Reachable from my laptop or my phone. That's what both sides of the argument are actually asking for.
On top of that, two modes of coding:
Exploratory - still figuring out the shape of the thing. Droid as the harness, driven through a modified Paseo build. That's also what lets me work from mobile, which matters most for long sessions that don't need much back and forth.
Mature - the product is cooked, now it's features and fixes. That runs on Cloop, our internal Devin-style environment, with a brainstorming module and Ladybug for hunting production bugs from discovery through to a PR. Human always in the loop at the PR.
Everything funnels into Droid underneath, so a session started from Paseo, from Cloop, or from MUXI is the same session.
The rest is deliberately boring. Apple Reminders for todos. Apple Notes for quick notes. Superhuman for email. Cloudflare, Hetzner and DigitalOcean for hosting. Slack Connect with clients, Discord internally. VS Code, which these days I barely use as a code editor - it's where I write books.
And I self-host almost everything I can. Not to save money. When you're spending tens of thousands on tokens, thirty dollars a month on a subscription isn't a line item worth thinking about. It's data ownership. I want to reach the database directly when I need to. I'll happily pay for your software. Just let me host it.
Thinking, not typing
Calls, clients, team syncs - all real, all less frequent than you'd guess.
Most of my time goes into thinking. And a lot of that thinking is about making the codebase smaller, because a smaller codebase is a smaller maintenance surface.
The exercise I keep running is first principles. Even on products that have been live for years with real users, I'll ask what the thing actually is and strip it to the core.
What is a coding agent? The Kanban boards and the workflows are secondary. It's a system that can code. Everything else is a feature you add back deliberately, one at a time, after you've earned it.
Which means a lot of my output is decisions about what not to build. I'll spend hours on that. Sometimes days. Occasionally weeks.
For the thinking itself, I use Claude to hypothesise and talk things through - not because it's smarter than GPT, but because I find it easier to talk to. Then I take the output to GPT and ask what it thinks. Feed that back to Claude. Go around again. Perplexity for research, where being current matters more than being clever.
I'm the human in the middle of that loop on purpose. It's fuzzy, it's the part I enjoy, and it's where judgment actually lives. I have no plans to automate it.
Generation is cheap. Judgment isn't.
I read every line the agents produce, and I comment on it heavily.
What I bring to the table isn't typing Python faster than someone else. It's seeing the problem in a codebase or a production environment before we get there. That comes from decades of getting it wrong in public and remembering how it felt.
Same reason my email is open exactly twice a day - once after the morning briefing, once at the end. There's a short VIP list that gets a fast reply, and it isn't ranked by importance. It's about not being the bottleneck. My accountant with something urgent uses a specific address. Routine bookkeeping goes in the normal pile.
The parts that aren't replicable
Two things here, and I want to be straight about both.
I don't really run Automaze. We're close to thirty people, and Anton and Alex run the agency - team, pipeline, builds. I win and keep customers, write the PRDs, and decide how products get built. Then the team takes over and I come back for the last ten percent. Lean founder doesn't mean solo founder.
Leads are 99% inbound, and that's not a strategy. People use the open source, trace it back, and reach out. The podcast. The newsletter. The books. None of it was built as lead gen. It accumulated organically over fifteen-plus years.
Trust me, when I released that code fifteen years ago I wasn't thinking about pipeline. So the unsatisfying part of the answer is that you can't run this next quarter. It's a byproduct of giving things away for a very long time, and the compounding only shows up late.
The FDE thing
Forward-deployed engineering is having a moment. An FDE is a customer-facing engineer embedded directly in the client's team.
At Automaze that's just what we've always done. What I do specifically is act as the FDE myself for every new client, and stay in the room for every new process we implement.
People ask whether that makes us look small.
I don't think it does. It's a design decision, and the reasoning is simple: I can't confidently talk about something I haven't experienced first-hand, and I can't improve a process I don't understand.
Small number of high-quality decisions. There's no high-quality anything without understanding first.
So I'm there at the start, my team takes over, and I come back at every fork in the road. Clients tend to like having the founder in the room working it out with them. For a company our size, I think that's the job.
Take the write-up
I put the whole thing into a PDF - the full setup, the tool list with links, the automation buckets, all of it. No email required, no gate.
I don't know if it's a blueprint you can copy. Probably not, and the inbound part definitely isn't. But the classification habit is, and it's the piece that actually did the work.
Pick something you've done more than three times this month. Decide which bucket it goes in.
That's the whole job.
Updated on 23 July 2026.
For the latest version and comments, please see:
https://aroussi.com/post/how-i-work
My job used to be doing the work.
Now it's deciding which bucket the work goes in.
That sounds like a promotion. It isn't. It's a different job with a different failure mode, and it took me longer than it should have to notice the switch had happened.
People keep asking how I get through it all - Automaze, the books, the podcast, the side projects, the stuff I build that never ships. Fair question. So here's the honest answer, including the parts that aren't replicable.
The whole system, in one line
I make 20 or 30 decisions a day and almost all of them are classifications.
Rails or agents. Bespoke or general. Human, automated, somewhere in between - or time to kill the process entirely.
That's it. That's the job.
Everything else you're about to read - the machine in my office, the harnesses, the briefings, the VIP list - exists to protect the context in my head so I can keep making those calls all day without the quality degrading. Bezos said he gets paid to make a small number of high-quality decisions. I'm not running Amazon. But that's the target, and everything else is scaffolding.
Automated doesn't mean AI
Here's the reframe most people need.
Anything I've done more than a few times, and have a process for, gets automated. But automation sorts into three buckets, and only one of them is agentic:
Deterministic. Known inputs, known outputs. No model anywhere near it. Just automation.
Hybrid. A fixed pipeline with one fuzzy step dropped in exactly where judgment is needed. Write the actual email. Make the call at this branch. Everything before and after stays dumb.
Agentic. Genuinely fuzzy end to end. An agent owns the whole task.
Most of what I run is the middle one.
The instinct to make the entire workflow intelligent is usually wrong. It's slower, it costs more, and it's much harder to debug. A deterministic pipeline with one smart step is easier to reason about than an agent that's smart the whole way through and unpredictable in fourteen places.
Boring where it can be boring. Smart where it has to be.
Why almost everything is gated
Most of my automations require human approval at some stage.
That's not caution about the technology. I'm all in on the technology - I spend tens of thousands of dollars a month on tokens. It's accuracy about the domain.
Every client we serve is an edge case. The processes look nearly identical between them and then diverge in exactly the place that matters.
So the dial between fully automated and human-gated is technically a setting, but in practice it's a judgment about how we want to work. AI takes the dirty work off my plate. It doesn't stand in for me in client communication.
The setup
Local or cloud is a real debate and both camps have a point. My answer is that my local environment is the cloud.
There's a dedicated machine in my home office, accessible to me and only me, from anywhere. Fully configured the way I want it. Safe to run experiments on. Reachable from my laptop or my phone. That's what both sides of the argument are actually asking for.
On top of that, two modes of coding:
Exploratory - still figuring out the shape of the thing. Droid as the harness, driven through a modified Paseo build. That's also what lets me work from mobile, which matters most for long sessions that don't need much back and forth.
Mature - the product is cooked, now it's features and fixes. That runs on Cloop, our internal Devin-style environment, with a brainstorming module and Ladybug for hunting production bugs from discovery through to a PR. Human always in the loop at the PR.
Everything funnels into Droid underneath, so a session started from Paseo, from Cloop, or from MUXI is the same session.
The rest is deliberately boring. Apple Reminders for todos. Apple Notes for quick notes. Superhuman for email. Cloudflare, Hetzner and DigitalOcean for hosting. Slack Connect with clients, Discord internally. VS Code, which these days I barely use as a code editor - it's where I write books.
And I self-host almost everything I can. Not to save money. When you're spending tens of thousands on tokens, thirty dollars a month on a subscription isn't a line item worth thinking about. It's data ownership. I want to reach the database directly when I need to. I'll happily pay for your software. Just let me host it.
Thinking, not typing
Calls, clients, team syncs - all real, all less frequent than you'd guess.
Most of my time goes into thinking. And a lot of that thinking is about making the codebase smaller, because a smaller codebase is a smaller maintenance surface.
The exercise I keep running is first principles. Even on products that have been live for years with real users, I'll ask what the thing actually is and strip it to the core.
What is a coding agent? The Kanban boards and the workflows are secondary. It's a system that can code. Everything else is a feature you add back deliberately, one at a time, after you've earned it.
Which means a lot of my output is decisions about what not to build. I'll spend hours on that. Sometimes days. Occasionally weeks.
For the thinking itself, I use Claude to hypothesise and talk things through - not because it's smarter than GPT, but because I find it easier to talk to. Then I take the output to GPT and ask what it thinks. Feed that back to Claude. Go around again. Perplexity for research, where being current matters more than being clever.
I'm the human in the middle of that loop on purpose. It's fuzzy, it's the part I enjoy, and it's where judgment actually lives. I have no plans to automate it.
Generation is cheap. Judgment isn't.
I read every line the agents produce, and I comment on it heavily.
What I bring to the table isn't typing Python faster than someone else. It's seeing the problem in a codebase or a production environment before we get there. That comes from decades of getting it wrong in public and remembering how it felt.
Same reason my email is open exactly twice a day - once after the morning briefing, once at the end. There's a short VIP list that gets a fast reply, and it isn't ranked by importance. It's about not being the bottleneck. My accountant with something urgent uses a specific address. Routine bookkeeping goes in the normal pile.
The parts that aren't replicable
Two things here, and I want to be straight about both.
I don't really run Automaze. We're close to thirty people, and Anton and Alex run the agency - team, pipeline, builds. I win and keep customers, write the PRDs, and decide how products get built. Then the team takes over and I come back for the last ten percent. Lean founder doesn't mean solo founder.
Leads are 99% inbound, and that's not a strategy. People use the open source, trace it back, and reach out. The podcast. The newsletter. The books. None of it was built as lead gen. It accumulated organically over fifteen-plus years.
Trust me, when I released that code fifteen years ago I wasn't thinking about pipeline. So the unsatisfying part of the answer is that you can't run this next quarter. It's a byproduct of giving things away for a very long time, and the compounding only shows up late.
The FDE thing
Forward-deployed engineering is having a moment. An FDE is a customer-facing engineer embedded directly in the client's team.
At Automaze that's just what we've always done. What I do specifically is act as the FDE myself for every new client, and stay in the room for every new process we implement.
People ask whether that makes us look small.
I don't think it does. It's a design decision, and the reasoning is simple: I can't confidently talk about something I haven't experienced first-hand, and I can't improve a process I don't understand.
Small number of high-quality decisions. There's no high-quality anything without understanding first.
So I'm there at the start, my team takes over, and I come back at every fork in the road. Clients tend to like having the founder in the room working it out with them. For a company our size, I think that's the job.
Take the write-up
I put the whole thing into a PDF - the full setup, the tool list with links, the automation buckets, all of it. No email required, no gate.
I don't know if it's a blueprint you can copy. Probably not, and the inbound part definitely isn't. But the classification habit is, and it's the piece that actually did the work.
Pick something you've done more than three times this month. Decide which bucket it goes in.
That's the whole job.
Updated on 23 July 2026.
For the latest version and comments, please see:
https://aroussi.com/post/how-i-work