Reviewing coding-agent output: attention is the bottleneck

Reviewing coding-agent output is the bottleneck now. A ping that a coding agent is stuck is no longer the scarce thing. The scarce thing is reading the work it says is finished. I put what finished, what is blocked, and what needs a yes into one brief, and I read it from 8:00 to 8:30 AM.

On 20 October 2026, Enzo wrote that three apps were on his feed that day, each one there to tell you which coding agent is stuck on a permission prompt. He split the job in two: "knowing its stuck was never the hard part. reading the 3 diffs it swears are done is." The next line is his: "my attention is the rate limit now."

The alert is getting built. The read is not.

What do the stuck-agent posts actually show?

The same morning, Manish Sharma posted Notch Island and called it a Dynamic Island for coding agents. The post says each agent gets a live card in the Mac notch: the prompt, the current tool, the last reply, and approve or reject with the diff. A click opens the terminal. The agent still picks the moment.

The day before, on 19 October, Denis posted Bellpip. The post says it puts a bell in the Mac notch, shows which session needs you, chimes, and a click lands on the tab. I am not counting that post as Enzo's "today." Same kind of tool, one day earlier.

Enzo said three apps and named none of them. Notch Island is named that same morning. Bellpip is named the day before. A name that is not in those posts stays off this page.

What they solve is "which one is waiting." What they leave is the second sentence. A card can put a diff in front of you. It cannot decide the diff is done. I wrote about owning that decision in who owns the architecture when AI writes the code. When several agents finish on their own clocks, the reads arrive in their order.

Does Anthropic call human attention the bottleneck?

How employment is changing at Anthropic, published 16 October 2026, says this acceleration makes human attention the bottleneck for agent output. The authors report that Claude sessions amount to around 80 percent idly waiting for human input, and that the vast majority of an interactive session is spent waiting rather than working. Even among non-technical employees, they write, reading all the output of their Claude agents would take half their work week. One engineer described the work as facing "an army of Claudes, with every one asking for my attention."

They describe the response as meta-work: visibility tools and monitors. A notch app tells you where to look. It does not do the looking.

How do you review coding-agent output in one window?

I keep one window for the brief. Agents write through the day. I do not open each diff when it lands, and a permission prompt is not a reason to switch tasks. I read it from 8:00 to 8:30 in the morning. A diff that waited overnight is the same diff at 8:00.

The window is the pattern I keep. It is not a timer I measured, and it is not one of the receipts below.

Northsun is invite-only. The names I use for this shape are Gargantua, Beacon and Reflector, and the names are not an offer.

The brief holds three kinds of line: a stuck prompt, a finished diff I can open, and a question. A still-running agent gets no line. A live card for every session would break the window.

A blocked agent stays blocked until that window. A permission prompt is not a person waiting on a call. Work that can wait, waits. Work that cannot wait does not belong in the batch.

An alert can say an agent stopped. It should land in the brief. The review stays in one window, with the diffs still attached to the lines.

FAQ

Should I approve a coding agent's prompt from the notification and keep going?

Approving a permission lets the agent continue, and it does not read the result. Notch Island's post puts the diff beside that step, closer to a review than a bell, and it still asks when the agent asks. The read wants a time that is yours.

Does a morning review window leave a stuck agent waiting all day?

A stuck agent becomes a line in the next brief, unless the work cannot wait. I am not reporting a measured count of minutes saved.

Is the 8:00 to 8:30 window in Anthropic's employment research?

No. Anthropic's 16 October 2026 note says human attention is the bottleneck, that interactive sessions spend most of their time waiting on a person, and that reading agent output can take half a non-technical work week. The 8:00 to 8:30 window is my pattern, and it is not one of their results.

Which apps were flagging a stuck coding agent on 20 October 2026?

Enzo wrote that three apps were on his feed on 20 October 2026, and he named none of them. Notch Island is a separate post from that same morning, and Bellpip is a post from 19 October. I am not naming an app those posts do not name.

The receipts

The posts and the paper are public. The morning window is my pattern, and it is not one of these receipts.