Your agents should live at home
My agents run on a Mac at home that they mostly set up themselves. Why that works for anyone who wants agents in their life, not just developers, and how to build your own.
AI narration. Sid is a clone of my own voice.
My agents don’t run on my laptop. They run on a 2021 MacBook Pro at home: plugged in, lid open, screen at zero brightness. I start work from my phone, check on it from my laptop at uni, and approve the next step from wherever I am. It runs day and night.
I barely touched it to set it up. I told agents what I wanted, and they installed the tools, wrote the background jobs and fixed what broke.
This isn’t really for developers. Most of them have their own version of this figured out already. It’s for anyone starting to use agents in their life: for professional work, for a business, alongside a 9 to 5, or to build a startup. It’s the journey of why this works, why how you think matters, and how to set it up.
One of the best ways to use this post is to paste it into your agent. It will read it and guide you through what you need to do. The technical details are folded away near the end, for you or for your agent.
A laptop is a bad place for an agent
A good agent task takes twenty minutes to an hour. It researches, writes, builds and checks its own work. Most of that time it doesn’t need you.
A laptop needs you all the time. Close the lid to catch a train and the work stops. It gets hot, the fans spin up, and every agent you start competes with you for the same machine.
Reading the code
So I moved them. The agents live on a machine that never moves, never sleeps and does nothing else. My laptop and phone just look in on them.
Why not the cloud
Most people assume the answer is a machine in the cloud. I don’t, for three reasons:
- It’s expensive. A capable cloud machine costs money every month, and a cloud Mac costs a lot more. I build iPhone apps, so I need macOS, and leaving my old Mac on costs almost nothing.
- It’s more painful to set up. A cloud box starts empty, and every tool, login and key has to be put there and kept there.
- My context lives here. This Mac holds my documents, notes, client folders and every project I work on. My agents read them to understand how I think and what each piece of work is for. At home, that’s just there.
Why T3 Code
T3 Code is the best agent orchestrator I’ve used, and the reason is how simple it keeps things. I know I can make things sound complicated. T3 doesn’t let me. You pick a project or a folder, start an agent in it, and when the work is done you settle the thread. The sidebar never fills up with a thousand old conversations, so I only see what needs me.
- Fix the flaky upload teststretch
- Review the pricing PRstretch
- Cut daily database I/Oargo
- Write the blog postblog
- Rename /used to /tech-stackrealsid08
- Clean up old worktreesmac
- Add the Runaway spreadrealsid08
- Update the résumécareer
- Ship build 48stretch
- Draft the supervisor emailscareer
6 settled
It runs on the Mac as one background service, and every other device is a window onto it: the desktop app, a browser tab, the phone app. Start a thread at your desk and it’s on your phone, mid-sentence. Close the app and the threads keep running.
the Mac at home · 2 watching
- Fix the flaky upload test
- Redesign the week screen
- Cut the planner’s cost per plan
It’s built with a lot of care by Theo, Julius, Maria and the people around them. I think it’s what agent orchestration will look like from here.
One bar at Lorne
Remote access is what makes all of this possible, and the T3 team clearly cares about getting it right. Some friends wanted to see the sunrise at Lorne, in Victoria. I had agents working on Stretch when we left, and I fell asleep on the way. When I woke up in Lorne, nothing on my phone would load: not Instagram, not Spotify, not even my WhatsApp messages. T3 Code loaded just fine, with every update from my agents. It felt a bit like magic.
- 2
I reach the Mac from anywhere, and it’s only open to my own devices. Early on, an agent stopped and asked me to click Allow on a dialog on the Mac. My reply, word for word: “its a REMOTE MACHINE IM NOWHERE NEAR IT”. So I set up a remote desktop I can open from my phone, then made sure I’d almost never need it.
No permission prompts
The biggest problem with running agents on macOS is permission prompts popping up at random. Can this app read your Documents folder? Can that one control your computer? An agent that hits one just waits, on a screen nobody is looking at. I had that for a good week.
So I gave the problem to an agent. It needed me a few times, and then it turned off almost every prompt for good. When something still needs a password, the agent asks me through T3 Code’s secure secret entry. I type it on my phone. The agent uses it through computer use without reading it, then deletes it.
- 1A dialog on the Mac asks for my password.
- 2The agent asks me for it through T3’s secret entry.
- 3I type it on my phone.
- 4Computer use types it into the dialog. The agent never reads it.
- 5The secret is deleted.
It has to look after itself
Left alone, it rots. After one week away I came back to simulators still running and dev servers nobody had stopped. Worse, T3’s worktree folder had grown to 191 GB, and only 45 of its 140 folders still belonged to a thread. The disk had 14.5 GB free.
- Worktrees
- Spotlight, indexing them
- Still running
Now it cleans up and updates itself on a schedule, and a reboot is a non-event. It still isn’t perfect. Just yesterday the disk filled up with eight agents running at once, and I had them clear caches before anything could carry on. Agents are also reluctant to delete things. They play it safe and refuse, which is usually right and sometimes maddening. I’m still working on a better answer, so if you have one, email me.
Not just for code
The agent has a whole machine, so it also has a browser and a computer it can use however it needs. It does what you would do on your computer, on your behalf and faster, and keeps you in the loop. You get more done, and spend less time at a computer.
Give it your context too. Connect your Gmail, your calendar and your Google Drive through each agent’s connectors. An agent that can read your inbox and your files understands what you’re actually working on.
Because the machine never turns off, it can take work without me starting it. For one client, a new issue or a production error starts a task on the Mac. An agent fixes it, sends the small fixes through with screenshots as evidence, and leaves bigger ones for my review. In its first three days it closed 39 issues.
Agents on this Mac also draft my emails, edit my videos and sort feedback for my app. One rule keeps that safe: they can read anything, but nothing goes out to a person until I approve it.
I think this is where work is going. Not one assistant in a chat box, but agents with a machine of their own, your context, and room to run many things at once.
The what and the why
I’m dictating this post right now. I’m talking to an agent through T3 Code on my phone, and it’s writing on the Mac at home, building the figures and showing me screenshots as it goes. That’s how I improve this setup, every other day, without spending much time on it.
Another thread is working on Stretch at the same time. I noticed a few cases in the app that I first thought were edge cases, then realised almost everyone would run into. So I said them into my phone, and an agent is now trying every flow on the Mac’s iPhone simulator. It understands what I want, how I think and how I like things.
- Me, dictated on my phone
If I ate something else, the price of whatever was planned should be taken off and the new price put in there.
- Agent
Reading the ate-something-else and eat-out code first to see how prices flow today.
- Agent
App shows the seeded week ($62.70 against a $60 budget). Recording flow 1.
- Agent
The pinned Save bar covers the price while typing. Since the price is the point of this change, moving it right under “What you had”.
- Agent
$62.70 − $4.50 + $14 = $72.20, correct.
That is the most important part of this whole setup. Every agent on my Mac reads one file that describes how I think, what I want and how I like to work, and I keep adding to it. Adapt your agents to you. Then they take care of the how and the small details, and you get to think about the what, and the why. The why always matters a lot.
Which subscriptions
If you’re hesitant to spend money on AI, I completely understand. Start with OpenCode Go. It costs US$10 a month and includes up to $100 a month of DeepSeek V4.1 Flash, which is kind of crazy. That model is close to the frontier, and plenty for setting up the machine, driving it and getting real work done.
For real products that go to production, pay for Claude Code or Codex. Claude Code is especially good at frontend and visual work, and Claude Opus is better at understanding what you mean. Codex is very good at computer use and at work that needs care and discipline. Try both. You don’t need the $200 plans: the $100 plans from Anthropic and OpenAI are more than enough.
Set up your own
You need a computer that can stay on: a Mac if you build for Apple platforms, and anything that runs Linux otherwise. You don’t have to do the rest by hand. Paste this page into your agent and it will guide you through the rest, one step at a time. There’s also a plain version written for agents, with every command and a check for each step.
Paste this into your agent
Set up this computer as an always-on home for my coding agents, using T3 Code. Follow https://blog.realsid08.me/agents-at-home.md step by step. Check that each step worked before you move on, tell me when you need me, and ask before anything you can't undo.
The detailsFor you, or for your agent. Tap to open.
The steps
- Keep it awake. Turn off sleep and have it restart on its own after a power cut. On a Mac, a launch agent running
caffeinate -dimsudoes it. - Install T3 Code and an agent.
curl -fsSL https://t3.codes/install.sh | sh, then sign in to an agent, for examplecodex loginorclaude auth login. - Run it as a service.
t3 upinstalls the background service and asks how your other devices reach it. - Pair your phone. Install the T3 Code app and scan the code that
t3 pairprints. - Hand the rest to an agent. Ask it to remove the permission prompts, schedule updates and cleanup, and write the two files below.
Getting in from anywhere
- T3 Connect is built in. It links up to 3 machines for free, or 10 with a paid T3 Chat plan. Phones and browsers don’t count.
- I use a Cloudflare Tunnel to reach T3 from anywhere, with no ports open on my router. Ask your agent to set it up.
- Cloudflare caps how big one upload can be, so some of my video files wouldn’t go through. For those I fall back to Tailscale, served over HTTPS from the Mac. Tailscale on its own is the easiest path if you’d rather not use Cloudflare.
- Either way, only my own devices can reach it. For the rare moment something needs a real screen, there’s a remote desktop over Tailscale.
Permissions
- Every agent runs in its full-access mode, so none of them stop to ask. Only do this on a machine that does nothing else.
- T3 Code has Full Disk Access, and a small job grants permissions again whenever an update moves an app.
- The prompts that are left go through computer use, with passwords from T3 Code’s secure secret entry.
- I turned off System Integrity Protection and Gatekeeper on this Mac. That’s a real trade. Don’t copy it onto a computer you can’t afford to lose.
One T3 server
- Only ever run one T3 server against one data folder. Early on I had the desktop app and the background service both open, threads froze, and the log said the database “already has an active writer”. Now the Mac runs T3 headless, and every other screen is a client.
Two files
~/AGENTS.md: how you think and how you like to work. The instruction files for Claude Code, Codex, OpenCode, Cursor and Gemini are all links to it, so one change reaches every agent.- A runbook: what runs on the machine, where it lives and how to fix it, written as the current state. Mine went from 600 lines of dated notes to under 200, and agents read it before they touch anything.
Upkeep
- Every day: clear old worktrees, build caches and simulators.
- Every hour: update T3 and the agents, but only when nothing is running. Everything else updates overnight.
- Check that the jobs actually ran. My updater once skipped every run for two weeks, politely waiting for five runs stuck as “queued”.
- Tell your agents once what cleanup may always delete, so a full disk doesn’t stall every agent at once.
Then close your laptop. The work keeps going.
Say hello
I love talking to people about AI, agents and this kind of tinkering. If you want to set this up and get stuck anywhere, message me on LinkedIn, X or by email. I’ll help you out. It’s easier than it looks.
I’m Sidhaarth Krishnan, a software engineer in Melbourne. More writing.