Rails

UpdatesNovember 2026

Many agents, one of you

My agents run in parallel. I run in series. How my days with them run, where the time goes, and what I changed.

On a typical day I have many threads of work open across my agents, and I move between them all day. Every one of those agents can work at the same moment as the others. I can't. There is one of me, so every thread waits its turn for my attention.

I keep a record of how I work with them, to see what that costs.

My days run in two shifts

Inside a working session I'm in near constant conversation with my agents. My messages come close together, one short note after another.

Most days run in two shifts. The first starts around 9 AM and runs to about 6 PM. After a break around dinner, a second shift runs from late evening to around midnight, quieter and with longer stretches on one piece of work. The early morning stays empty.

Most of what I send is short, a sentence or two. I'm steering, mostly with a yes or a pick between two options. Most days have a few bursts, when several agents come back at once and I answer them in a row.

The agents answer fast, usually within a minute, so the waiting in the loop happens on my side.

midnightnoonday shiftnight shiftday shift, about 9 AM to 6 PMproduct and recruiting work,built and reviewed with agentsnight shift, late eveningto around midnightwriting, and lining up whatthe agents carry overnight
Most days run in two shifts: a long one from about 9 AM to 6 PM and a quieter one from late evening to around midnight.

Each agent carries a whole area of work

An agent on Rails has four parts: a playbook that says how it works, a memory, the tools it can use, and its own app. Chat is the host where I talk to it. Agents like these carry most of my working life, with product, recruiting and writing among them.

Capability is abundant now. Attention is the scarce input, and the agents all draw on the same supply of it.

On a typical day I talk to several agents across several threads, and most of what I type goes to an agent's main chat.

Every switch costs me more than it costs them

Often my next message goes to a different agent or a different thread than the last one.

Some days stay on one track and some scatter. A quiet day can stay with one agent from start to finish. A busy one moves across many agents, and the switches add up fast.

A quiet dayone agent from start to finishA busy daythe color changes often, across many agentsEach color is one agent; each break is a switch to another.
Some days stay with one agent. Busy days move across many.

A switch is cheap for an agent. For a person it is expensive, and the research on that goes back twenty years. Gloria Mark, Victor Gonzalez and Justin Harris watched people at work and found that when a task was interrupted, it took about 25 minutes to get back to it, with two other activities in between. Erik Altmann and his colleagues found that an interruption of 2.8 seconds doubled the error rate on the task people came back to. Sophie Leroy named what's left behind attention residue: part of your mind stays on the unfinished thing, and the next task suffers for it.

There is a queueing reason too. John Little's law says the time each item spends in a system equals the work in progress divided by the rate it gets finished. The agents raise the work in progress almost without limit. The rate it gets finished is capped by the one person who has to read it, decide and say yes. More open threads means each one waits longer for me.

It is very cheap to start 100 tasks and very expensive to close 100. Agents made the starting free, and the closing still runs through me.

Most of the work is rereading

My agents run every day, and most of the tokens they process are cached rereads of earlier context. Only a small share is new input and output.

So more than nine of every ten tokens are an agent reading back what it already knew before it does the next step. That rereading is the part I'd otherwise have to do in my head every time I came back to a thread. The agents carry the context so I don't have to.

rereadcached rereads of earlier contextnewfresh input and output
More than nine of every ten tokens my agents process are rereads of earlier context.

If the agents hold the context, my job is to come back to as few threads as possible, and only when each one is ready for me.

What I changed

I now run my day in three modes.

Nicholas Fitz, Kostadin Kushlev and their colleagues ran a study with 237 people where checking notifications three times a day beat both constant alerts and none at all. Piotr Adamczyk and Brian Bailey found that interrupting at a good moment instead of a bad one cut annoyance by 56%. Paul Graham made the same point for makers years ago in Maker's Schedule, Manager's Schedule: a meeting in the middle of a maker's afternoon breaks the whole afternoon.

A few rules hold the modes together:

The goal is to have nothing waiting on me, with anything that needs me at the top. I call that agent zero.