The art of working on the go

On October 27 at 8:57 am, I was away from my desk. Overnight, a colleague had reported that a page on one of our company sites was coming up blank.

From my phone, I texted the agent that manages my other agents: "Yo what happened."

Nine minutes later an agent had the cause: a timing bug, two pieces of code racing each other as the page loaded. Ten minutes after that it had a fix, plus tests that proved the old code broke and the new code didn't, and it asked for my OK to put it live. I read it over and sent "Ok to deploy." Five minutes later the fix was live, and the agent had loaded the page five times in my signed-in browser to prove it. Then I sent "post it," and the explanation went to the team's thread.

From my first text to a checked, live fix took 46 minutes, and 21 of them were me reading. I sent eight words in all. My longest message was three words, and one of them was "Yo."

• 8words

  to get a broken page found, fixed and checked live

• 2min

  to unblock a waiting agent from my phone

• 7s

  for an agent to start on a phone message

• 318

  messages to agents from my phone, on 30 of 40 days

A TV remote doesn't play the show. The TV does. My setup works the same way: my Mac does the work, keeps the files and holds every sign-in. My phone only steers.

An agent, here, is an AI that does the work itself: it runs programs, edits files and checks its results. I keep about twenty of them busy on my Mac. The analogy breaks in one useful way: this remote works from anywhere, as long as the TV stays on. My phone reaches the Mac three ways:

• The Claude app shows every agent running in Claude Code, Anthropic's coding agent, on my Mac. I can read, reply, approve and start new ones.
• The ChatGPT app lets me answer and steer OpenAI's Codex agents on the Mac.
• Slack is where I text my browser agent, Aside, which works in my signed-in browser on the Mac.

Anthropic describes its version plainly: "Claude keeps running locally the entire time, so your code execution and filesystem access stay on your machine."¹ Nothing heavy runs on the phone. It only carries my messages.

It's easy to assume the phone is a lesser way to work, good for quick yeses and status checks, with the real work saved for the desk. So I compared 318 messages I sent from my phone with a random 220 I typed at the Mac.

The work is the same. On both, a third or more of my messages make a decision, about a quarter say go, and a fifth to a quarter ask where things stand. The messages are the same length, a median of 15 words. And agents start on them just as fast: 6.6 seconds after a phone message, 7.2 after one from the Mac.

Two things differ: the hours, and the errands. Two thirds of my phone messages go out before 9, after 6 or on a weekend, against a third of my Mac messages. And 18% of my phone messages are life errands, like a flight check-in or a bill, against 8% at the desk. The same phone that ships fixes to customers also checks me into a flight.

The real payoff is speed when an agent gets stuck.

Agents ask me things all day: which option, is this OK to ship, did I mean this or that. When an agent asked me something while I was away from my desk, I answered from my phone about half the time, in a median of 2.2 minutes. That's about as fast as I answer at my desk. The other half waited until I sat back down: a median of 3.4 hours.

Three hours is a long time for a job to wait for one person to sit down. I wrote this to an agent on October 19: "I'm at my phone. Surely there's something you could do."

Most of what I send away from my desk is small. Some of it isn't. Production, here, means the live system that customers use.

• A key replaced in under three minutes. On October 17 I sent 17 words from the ChatGPT app: another agent needed a production key for our email provider replaced. Two minutes and 54 seconds later the new key worked, and the agent had proven the old one was dead.
• A nightly job, first fix in eight minutes. One evening, a nightly job at my day job failed. I started it with one sentence from my phone: "Is this system broken? Was there a hard failure? Fix this." Eight minutes and a few short replies later, the first fix was live. The next morning, the agent had built guards so it couldn't fail the same way.
• 75 descriptions at 12:50 am. One night I sent my browser agent a 21-word Slack message: there's Latin filler text on one of our sites. Twelve minutes later it had fixed the page and all 75 catalog descriptions, published them and checked the live pages.

Early on, the phone kept getting in the way. Agents wrote me essays. A yes I gave in one thread didn't count in another. And some jobs needed the desk, like a confirmation email I couldn't open on my phone.

Each time, I changed the system instead of my habits. On October 9 I wrote to an agent: "So I'm currently on my phone, and you just gave me a ton to read. Can you just boil it down to core bullet points of what I need to know or what I need to decide on right now?" Messages like that became rules. My status report now leads with what needs me, numbered, so I can reply "1 yes, 2 B" with my thumb.

The same thinking runs through the rest:

• A yes I type in any agent's thread counts in every thread, after a script finds my exact words.
• One status page shows what needs me first, and it fits a phone screen.
• Every reply ends with the decisions at the bottom, and buttons that answer them.

The confirmation email became a rule too: anything that needs my signed-in browser goes to my browser agent on the Mac. Teach your AI once is about that habit.

A remote control only works if the TV is on. My Mac stays plugged in, and a small program keeps it from sleeping.

I saw what happens otherwise in late October. For a few days the Mac ran unplugged with the lid shut, and it slept up to 4.8 hours a day. For those hours my phone had nothing to talk to. OpenAI's docs say it plainly: putting the computer to sleep stops remote access.² Claude's version reconnects once the Mac wakes up, which helps only after it does.¹

The obvious worry is that this turns life into one long workday. My data can't settle that, but it does say the phone didn't shift my work toward the evenings.

In the 11 weeks before I set this up, 50% of my finished code changes landed outside weekday 9 to 6. In the 6 weeks since, 45%. The total went up both inside and outside office hours, mostly from other changes I made that same week, so I can't claim the phone kept my evenings free. What it changed is how long a stuck job waits for me.

Studies of after-hours work link burnout and poor recovery to two things: feeling pressure to answer right away, and never switching off.³⁴ My agents don't pressure me. They wait, and they keep working on everything that doesn't need me.

Attention on the move comes in bursts. In a 2005 study, people searching the web on phones in busy city settings went from attention spans of over 16 seconds in the lab to bursts of a few seconds.⁵ That's why my agents lead with the decision, and why a reply can be one word.

Talking beats typing on a phone. In a 2017 lab test with short messages, speech was about three times faster than an iPhone keyboard, 153 words a minute against 52.⁶

And small decisions fit small moments. Researchers at Microsoft and Stanford found that work broken into small pieces took longer overall, but came out better and may hold up better to interruptions.⁷ A decision, a yes or a question is a small piece. The long write-ups wait for the desk.

• My logs don't record which device I used. "From my phone" means the phone apps, minutes when nobody touched the Mac, or my own words saying so, so the counts are a floor.
• I answer the easy questions from my phone, the ones that take a few words, so the two-minute figure flatters the phone a little. It covers Claude Code agents only.
• I sorted the messages myself, in one pass. The phone-and-desk comparisons come from the same weeks; the merge comparison spans a month when several things changed at once.

1. Keep the work on one machine that stays on. Let the phone steer it, not do it.
2. Make every answer phone-readable. Tell your agents to lead with the decision and number their questions, so you can reply with one word.
3. Let a yes travel. Approve once, from anywhere, and let the agents check it's really you.
4. Hand signed-in errands to a browser agent. The phone can't log in everywhere. The Mac already is.
5. Use the phone to unblock, not to stay on. Answer what's stuck, then put it down.

Working from my phone is mostly steering from farther away. The art of steering an agent is how I talk to agents mid-job, the signed-in browser is how they work behind my logins, and Helm is how I keep track of it all.