What to remember from this issue
Agents do not get faster inside processes built for people. Find what each step assumes about a human and rebuild that step, but keep the step where someone has to answer for the result.
The field note
This month Anthropic explained why it stopped using Git, the tool almost every software team uses to keep track of its code. In May the company merged about 2,500 code changes a day. By September it merged more than 30,000, and nearly all its users are now agents and automated jobs, which never go offline.
Git was not broken. It was built around a person who works on a laptop, sometimes without a connection, and needs a full copy of everything. In the team’s words, “Git assumes that clients work offline and need a complete copy of the repository.” An agent is always online and needs one file at a time. So a team of seven people built a replacement and moved the company’s main code repository onto it in five months.
The same thing happened to code review. Nobody reads 30,000 changes a day line by line, so reviewers “now examine the intent of the change as opposed to reading every line of code.” The step stayed, but what it assumed about a human (how fast a person reads) went.
This is the gain most large companies are not getting. They put agents inside processes built for human speed: the weekly approval meeting, the form that waits for a signature, the report nobody opens before Thursday. The agent finishes in a minute and then waits a week. Two things block the fix. Nobody sees the human assumption inside the step, because it has always been there, and when someone does see it, the organisation protects it.
Not every slow step is waste, though. Some steps are slow because someone has to answer for the result. Anthropic kept that part: the team still reviews every change, and in their words “the role of the engineers was to tell Claude ‘no.’” Automate the waiting, keep the signature, and make the signature fast.
Try this week
Take one process your team runs every week and write each step on its own line. Next to each one, write what it assumes about a person: that someone reads it, that someone is at their desk, that someone replies within a day. Circle the steps where a person has to answer for the result. Every step you did not circle can be rebuilt for a machine.
Seen this week
A shopping agent decided overnight which shops existed. Around 10 July, ChatGPT’s shopping mode started reading merchants’ product files instead of their web pages, and the share of results drawn from those files rose from 8.26% to 61.54%. Profound, a firm that sells visibility tools, measured this on 687 of its own customers: 450 lost a third of their visibility or more. OpenAI’s specification is blunt, because a product without a valid stock status is rejected. Shops whose catalogue was written only for human eyes were gone the next morning. Profound, “How ChatGPT 5.6 transformed Shopping mode” and OpenAI’s product feed specification
Someone said it better
“coding is automated, code review becomes a bottleneck. Code review is automated, judgment on what to build and what not to build becomes a bottleneck.”
Jeetu Patel, President and Chief Product Officer of Cisco, which has 32,000 engineers, on Y Combinator, 8 October 2026, at 9:50. Anthropic’s story is his first sentence. His second sentence is where the work goes next.
One number
19%: the employment gap for young workers in the jobs most exposed to AI, measured by Stanford’s Digital Economy Lab on ADP payroll data. The same page says there is “No Widespread Displacement”. Stanford Digital Economy Lab, Canaries dashboard, last updated 21 October 2026