Yesterday, I merged 109 code changes, or pull requests (PRs), at Metaview. It was a heads-down day, broken up by lunch, a bike ride and a few strong cups of tea. However, when the 100th PR went in at about 6pm, I paused a little bit and asked myself, ‘What is my job?’

When I started coding professionally, I'd have been stretched to do a PR a day. And although PRs are an awful measure of productivity (almost up there with counting the emails someone sends), going from one a day to 109 in a day says something. I'd love to put that change down to my own competence. But I’ve not become a 100x engineer.

Most teams have bolted AI coding assistants onto pre-existing software lifecycle processes. The AI assistant helps an engineer write a function, a reviewer gets to it at some point, and it ships in next week's release. This way of working was fine when writing code was the bottleneck in software factories. But now that an agent can autonomously write the change in minutes, the bottleneck has been lifted from coding, and we need to reorganise the factory as the bottleneck moves to reviewing, running tests and deploying.

At Metaview we've spent the last year chasing these bottlenecks. Getting to 109 PRs didn't mean working 996. It meant smart engineering. Deploys (and rollbacks) are instantaneous, the test suite runs across 16 machines in two minutes and AI reviewers read every change, so only the high-risk ones wait for a human. Once a change is live, we constantly compare its error rates with normal levels to monitor for issues early.

It all runs on Anthropic's Opus 5.5 in fast mode, inside an internal agent built by a colleague (kudos to Keir) and integrated into Slack. Every Slack thread with our agent gets its own cloud computer and its own copy of the code. The agent then writes the change, watches the tests, argues with the review bots and checks the code in production once it's live. And because it all lives in Slack, I can check in on the agents from the bus or out on a bike. It’s worth noting that in the whole day I didn’t open an editor like Cursor or VS Code, just Slack and GitHub (which made it through the day without an outage…).

Plumbing alone didn’t get me to 100, though. I was working off a detailed spec (credit to James) that our engineering team had argued over until it contained firm opinions an agent could root itself in. Alongside this, the work was verifiable. The agent had an opinion on how to build it and had a way to check if it was right or wrong, fix the code and repeat. My job became reading the results and deciding what came next.

This new way of working is weird. A few years ago, a good day was one change I'd written myself in VS Code. Yesterday I spent most of the day reading what agents had done in Slack and telling them what to do next. If every engineer on a team had a 100 PR day, I’m genuinely not sure how that team would work. How do you roadmap that much output? Do MVPs still make sense? How does a PM, or a sales team, keep up? Plenty of people claim to know what this future looks like. Truth be told, I’m not sure anyone is ready for it. It was my first 100 PR day, but I doubt it'll be my last.

Originally published on Casa Mayhew on 5 November 2026.