Shipping From Bed: Why I Run Cursor Agents From My Phone Now

Ed Stevenson·May 13, 2026

It is just past eleven. My laptop is closed on the desk in the other room. I am horizontal, the lamp is dim, and the only light on my face is a phone screen running Cursor.

I type a few sentences — something like "add a blog index page and a dynamic route for individual posts, match the rest of the dark theme, write a first article" — and tap send. A cloud agent picks it up, branches off main, reads the codebase, makes the edits, runs the typechecker, pushes a branch, and opens a draft PR.

I have not touched a keyboard.

This used to feel like cheating. Now it just feels like the new shape of the work.

The shift: from "writing code" to "directing agents"

For a long time, the unit of progress in software was a keystroke. You opened the editor, you wrote a function, you ran the tests, you committed. The loop only moved when you moved.

Agents change the unit of progress. The new unit is intent. You describe what you want — clearly, concretely, with the right context — and an agent goes off and converges on it. You review. You correct. You merge.

A few things follow from this shift that I think people underrate:

Latency stops being your bottleneck. You no longer need a perfectly set up dev environment in front of you to make progress. You need a clear thought and a way to send it.
You can run work in parallel. While one agent is refactoring your blog, another can be writing tests, and a third can be drafting a migration. They are background processes for your codebase the same way your subconscious is for your brain.
The skill stack changes. Knowing how to phrase a task, what context to attach, how to verify the result, and when to throw it away and start again becomes more valuable than knowing the exact git incantation.

This is not a "the AI does everything" essay. Reviewing matters more than ever. Security matters more than ever. Taste matters more than ever. But the part where you sit in a chair and type the boilerplate? That part is dissolving.

Why the phone, why bed

I want to be careful here, because I know how this sounds. "Bro, just code in bed!" is dystopia-coded if you take it the wrong way. So let me say what I actually mean.

For most of my career, the laptop has been a small gravity well. To do anything real, I had to come back to it: sit at the desk, plug in, open the IDE, get warm, get going. That gravity well is great for deep work, but it is terrible for the little ideas that show up at the wrong time.

The shower idea. The on-a-walk idea. The "wait, the onboarding flow is broken on Android" idea I have right before falling asleep. Historically those got jotted into a notes app and… mostly died there. Some of them got rediscovered weeks later, half-remembered, no longer urgent.

A phone-driven coding agent fixes that, gently. Lying in bed last night, I noticed my own site was advertising a /blog that just said "Coming soon." I'd known about it for ages. The cost of fixing it had always been "wake the laptop, swap context, do a thing." Last night the cost was: pick up the phone, describe the change, tap send, sleep on it, review in the morning.

That is the actual pitch. Not "work more, work in bed." More like: let the small, persistent ideas survive long enough to ship.

A field manual for mobile-agent workflows

Some practical things I've learned doing this for real, not as a stunt:

Write the brief like you're texting a sharp junior engineer. Be specific about what and why, not how. "Add a blog list page that reads from _posts/, match the existing dark theme used on projects.tsx" beats "make the blog work."
Point at the right files. A two-line message that says "see pages/projects.tsx for style, see lib/api.ts for how posts are loaded" produces dramatically better output than a paragraph of vague vibes.
Always ask for a PR, never a direct push. This is non-negotiable. Agents are great. They are also overconfident. A draft PR with a diff you can scroll through on your phone is the smallest possible safety net, and it is enough.
Defer review, don't skip it. Bed is great for kicking off work. It is a terrible place to merge. I keep a rule: nothing an agent produced at night lands on main until I've read the diff at a desk with coffee.
Treat secrets like a phone is a public place. Don't paste API keys into chat. Don't approve workflows that exfiltrate credentials. Read the proposed changes for anything network-shaped before you click merge. Agents will happily add a dependency you didn't ask for; you have to notice.
Keep the loop tight. One small task per agent run. "Add the blog index" is a good task. "Rebuild my entire site" is not. Small tasks fail safely. Big tasks fail expensively.

What this is good at, what it isn't

It is good at:

Wiring up the obvious plumbing on a project you already understand.
Knocking out the boring 80% of a feature so you can spend your real focus on the interesting 20%.
Keeping forward motion on a side project that would otherwise rot in a branch named wip-finally-fix-this.
Triaging "is this fixable in ten minutes?" without leaving the couch.

It is not good at:

Making product decisions for you.
Replacing a thoughtful review.
Understanding constraints that live entirely in your head and were never written down.
Anything where being wrong is expensive and irreversible — migrations, destructive scripts, anything touching prod data.

The mental model I keep coming back to: an agent is a very fast intern with infinite patience and zero context outside what you give it. Treat it accordingly.

The bigger picture

I think we're going to look back on this period as the moment where "writing software" quietly stopped requiring "being at the computer." Not because the computer disappeared, but because the interface to the computer did. A phone, a voice, a description — that's now enough to move the needle on real projects.

I run Konuke around the idea that most teams are still doing by hand what could be running on autopilot. Agents on your phone are the personal version of the same insight. The work that used to live in your head as "I should really get around to that" can now actually leave your head and become a pull request, while you sleep.

This blog post itself? Kicked off from bed. Reviewed at a desk. Merged in the morning.

That feels like the right loop.

If you want to try this: install Cursor, set up cloud agents on a real repo (start with a personal site or a side project, not something with prod traffic), and the next time you have a small "I should fix that" thought at 11pm — instead of letting it die in a notes app, send it to an agent and go to sleep.

Curious what else this workflow makes possible? More notes from this experiment are on the way. In the meantime, there's more on automating the unglamorous parts of work over at konuke.com.