Rails

UpdatesNovember 2026

Agent zero

Running many agents gives you power and wears you down for the same reason. Agent zero is inbox zero for agent work.

Most days I work with my agents in two shifts, a long one through the working day and a second one late at night. Many threads of work move at once across product, recruiting and writing, and I move between them all day.

A typical day gets a lot done. Agents write code, draft pages, find candidates and keep research moving while I'm busy with something else. It also takes a lot of my attention. I've come to think both facts have one cause.

Every agent is one more thing that can ask for you

An agent is useful because it works without you. You hand it a job, it runs, and at some point it comes back. It finishes, gets stuck, needs a decision, wants an approval or finds something you should see. Each of those is a moment where it asks for you.

Add an agent and you add work done in parallel. You also add one more thing that can pull you out of whatever you were doing. The power and the drain grow together, because more agents working means more agents with a reason to reach you.

1 agentyoua few agentsyoumany agentsyouasks coming back to youwork going out
Every agent that does work for you is also one more thing that can ask for you.

Most advice treats these as two problems. One is getting more out of agents, and the other is protecting your focus. I see one problem. Capability got cheap. My attention stayed as scarce as it always was, and every agent spends some of it.

A typical day runs in two shifts

I keep a record of how I work with my agents. Most days run in two shifts, a long one from about 9 AM to 6 PM and a second one from late evening until after midnight. Inside a working session I'm in near constant conversation with them, mostly short replies: a yes, a pick between two options, a note on where to go next.

The day shift is the dense one. Late morning is usually the busiest stretch and the afternoon runs steadier. The night shift is quieter, with longer stretches on one piece of work. Through both, product, recruiting and writing take turns in front of me.

12 AM6 AM12 PM6 PM12 AMday shift, about 9 AM to 6 PMproduct and recruiting work,built and reviewed with agentsnight shift,late evening topast midnightwriting, and lining up whatthe agents carry overnight
A typical day with my agents: a long day shift and a quieter night shift.

A year ago I could not have kept that many projects moving by myself. The cost shows up in the switches. Many times a day I leave one thread of work and pick up another, often in the middle of a thought.

The toll comes in through the same door

The cost of a switch was measured long before agents. In 2005, Gloria Mark, Victor González and Justin Harris followed information workers through their days. More than half of their work segments were interrupted, and people who came back to an interrupted task the same day took about 25 minutes to get back to it, with about two other activities in between.

Six years later, Laura Dabbish, Mark and González found that about half of all interruptions are self-made, and that being interrupted from outside makes people more likely to interrupt themselves in the hour after. That matches how agents feel to me. Once a few of them have pulled me out of a task, I start checking on the rest without being asked.

Sophie Leroy showed in 2009 that when you leave a task unfinished, part of your attention stays with it, and you do worse on the next one. With many agents, almost everything is unfinished. Every open thread holds a little of my attention even while it sits quiet.

How fast it all feels hides some of this. In a 2025 study by METR, experienced developers working in their own projects took 19% longer on tasks where they could use AI. Afterward they still believed AI had made them about 20% faster. Starting something with an agent takes seconds, so the work feels quick. The time goes into what comes after the start, the prompting, the waiting and the reviewing.

Approvals cost a wait

Everyone who puts agents on real work puts some approvals in front of them, and they should. An agent that sends email or changes production data ought to check with a person first.

Each approval is also a point where the agent stops and asks for you. One costs nothing. Across many agents they become a steady stream of small asks through the day. Each one either breaks your focus now or leaves an agent idle until you get to it.

When the ask arrives changes what it costs. In 2004, Piotr Adamczyk and Brian Bailey found that interrupting people at a good moment in a task, instead of a bad one, cut their annoyance by 56% and their frustration by 49%. Paul Graham made the same point about calendars in his essay on the maker's schedule. For someone who makes things, one meeting in the middle of the afternoon can cost the whole afternoon.

So for every approval I want two answers. Where should it wait for me, and how much can the agent settle before it asks? An agent that brings a finished recommendation and needs a yes costs me seconds. An agent that brings a question it could have answered costs me a switch.

Cheap to start, expensive to close

The other pattern every heavy agent user meets is the open loop. Starting work with an agent takes a sentence. Closing it means reading what came back, deciding whether it is right, checking that it works and marking it done. You can start a hundred tasks in an hour. Closing a hundred takes days.

Agents widen that gap. They start things eagerly and rarely finish the last step on their own, so a draft waits for review or a finished change waits to be merged. Each one keeps a small claim on my attention, the way Leroy's unfinished tasks do.

Closing has to be part of the agent's job. An agent that opens a loop should own closing it, and should bring it to me only when the last step is mine to take.

Agent zero

I call the goal agent zero. It is inbox zero for agent work. Nothing waits on you, and anything that does need you sits at the top.

Inbox zero was always about every message being dealt with, answered or filed or deleted, so the inbox stopped holding claims on you. Agent zero applies that to threads of work. The agents keep running. What goes to zero is the count of things waiting on me that I haven't seen.

Focus by default, an inbox a few times a day

Getting there changes how the day is built. I keep every agent and cut the ways they can reach me down to a few.

Focus is the default, with one thread of work and one conversation on the screen and nothing else. Simon Willison, who writes one of the most useful blogs on working with coding agents, put the limit plainly: "I can only focus on reviewing and landing one significant change at a time." Agents run in parallel. I run one thing at a time.

Everything else lands in an inbox. It opens a few times a day, and I work it one item at a time, with that item's context pulled up beside it so I never have to go find the thread it came from. In 2019, Nicholas Fitz and colleagues tested this pattern on 237 people's phone notifications. People who got their notifications in batches three times a day came out ahead of people who got them as they arrived, and people who got none at all grew more anxious.

Only true urgency breaks in, and the bar is high. It means a running agent stuck on something only I can do, where waiting costs more than the switch. Even then it should wait for a natural pause in my work if it can, which is the Adamczyk and Bailey result put to use.

12 AM6 AM12 PM6 PM12 AMfocusfocusfocusagents keep workingurgent waits for a pauseinboxinbox
The day I'm working toward, on the same clock as a typical day.

The day I want keeps the same two shifts and runs on Graham's maker's schedule, with long blocks of focus and a few short passes through the inbox.

What I'm building

I'm building Rails, a managed agents platform, to make this the normal way to work with many agents. Agents bring finished work and clear asks. Everything that needs you lands in one place, with the most important item first and its context open beside it. The agent that opened a loop is the one that closes it, and the agents keep working while you focus.