I have the brain for it, and I have realized that in this agentic world, it's time to multi-task. Fortunately, though my brain feels a bit jumbled at the end of a work session, I've found that working asyncronously and managing my own mental context easily has been the best thing to focus on.
I've built a few apps lately that have been fun and served their own needs to satisfy my itch in product work, and I realized that I wanted a better workflow to work on a few of them at a time.
For a while my "multi-project" workflow was just a lot of terminal tabs. One claude running in each, one project per tab, and me CMD > swapping them trying to remember which one I'd left mid-thought. Start a refactor in one, kick off a migration in another, tab back to the first. The pane is too filled with text. Resize it so I can read it...the disaster goes on. Absolute dumpster fire.
The part that actually hurt was how fragile it all was. Close the terminal by accident and the session is gone. Reboot and they're all gone. Step away from my desk and I'm at lunch with an idea? Too bad. Lost. An agent would finish, sit there waiting for input, and rot for an hour because I was in a meeting two rooms away.
I could just start a single Claude Remote Control session or clear a few I've started, but this seemed better and more organized. I love AOE because it allows me to save all of my sessions and archive them, too, so I don't like clearing a lot.
Fresh sessions = less context = fresh start = better - me 😝
Three pieces fixed this. Each part serves its own wonderful purpose, and together they changed how many things I can have going at once and really how much work I can do.
1. Sessions that survive the terminal
The base layer is a tool I run called aoe (Agent of Empires) — a session manager for coding agents. Instead of claude living and dying with a terminal tab, each project gets its own persistent session backed by tmux. Close iTerm, reboot the Mac, come back the next morning — the session is still there, the agent process is still there, and it remembers where it left off.
There's a dashboard that lists every session at once, so "which projects have something running" is a glance instead of a tab hunt.
Demo GIF courtesy of the Agent of Empires team.
That alone took the fear out of running more than one thing. If a session can't disappear on me, I'll happily have four of them going.
2. Steering them from my phone
The next problem was being tied to my desk. aoe serves that dashboard as a local web app, so I put it behind a Tailscale Funnel — a stable public URL, gated by a passphrase, that installs on my phone as a PWA and is only accessible via VPN. Now the full dashboard, every session, live terminals and all, is one tap away on my phone.
So the flow becomes: kick off a few agents before I leave the desk, then approve prompts, answer questions, and redirect them from the couch or the coffee shop. It's the same session I started on my laptop, not a read-only status page — I can actually type into the terminal.
Sometimes, I noticed the server would go randomnly down and I'm working on a more stable solution, but for now there's a tiny launchd job that checks every two minutes and bounces the tunnel if the public side is down but the local side is fine.
3. Starting work with a text message
The dashboard is great for steering sessions that already exist. But one thing was still clunky: starting new work when I'm away from the desk. Spinning up a session from a phone browser is fiddly — find the project, open a session, type into a cramped terminal. That friction is exactly where a good idea dies before I get home.
So the last piece is Telegram. aoe has a conductor — one long-lived Claude session I've paired to a bot. I don't code in it; its one job is to start things. I text the bot "start a session on the portfolio repo and update the blog copy," and it spins up a fresh child session, points it at the right project, and kicks it off. No laptop, no terminal, no setup.
I needed a way to tell Telegram about my current projects. Read on, my friend.
How the Telegram bot knows which repo to use
The bot maps a #hashtag in a message to an absolute repo path on my computer. It lives in a repos.json that I edit freely — alias on the left, folder on the right:
{
"myne": "/Users/andrewmartin/code/myne-app",
"releasewave": "/Users/andrewmartin/code/releasewave",
"rw": "/Users/andrewmartin/code/releasewave"
}Aliases are cheap, so a repo can have a few (rw and releasewave both land in the same place). I text something in Telegram like #rw the release page is 404ing and the bot pulls the first tag that matches a key, resolves it to that path, strips the tag out of the message, and uses what's left as the actual prompt for the new Claude session. I could paste that code here, but it's not that interesting or novel. The idea is.
Now, I can type in something like
#rw validate security on trpc endpointsHere's a few pictures to see it in action:
Here's a little sample of it in action:
Resuming on Claude Remote Control
Telegram is only the ignition, though. Once the session's running I don't try to have a real conversation over chat — I switch to Claude remote control from the last section and drive it there, where I can actually see the terminal and type into it properly. Text to start, remote control to work. The bot just gets the engine turning from wherever I am.
And finally, back in Claude's official app:
I can also stop and restart the sessions from Telegram, if something goes wonky. Beautiful!
What changed
I can run three or four projects at once now and not feel like I'm dropping any of them. The sessions can't vanish, I can start one with a text from anywhere, and I steer them all from the same dashboard whether I'm at my desk or not.
The thing I'd tell anyone drowning in agent tabs: the win isn't always just a smarter model. It's also making sessions durable, reachable, and cheap to start. Once kicking off a fifth project is one text message and none of them can quietly die on you, running several at once stops being stressful and starts being the point.
The Future: Harnesses and Claude
After going to AI Engineer World's Fair and sitting in on the Anthropic session on harnesses, I'm going to work on building a Mac Mini and drive more powerful harnesses around the containers AOE runs in. But that's for another day.