The 11x Engineer

March 10, 2026

This month I'm launching my next big solo dev project. It has close to 1700 commits, and everything is authored by me in git history, but the latter half of these commits could hardly be called written by me at all. All of the real work done was done by agentic AI.

Like many software engineers, in 2025 I gradually started adopting an AI agentic Codex. As my usage ramped up I added Anthropic's Claude Code as well, both on their $20/month tiers, and the more I used them, the faster things moved. I got curious exactly how much success I was having, so I asked the more verbose one - Claude - and it came up with a deep dive.

The conclusion Claude drew is basically mine as well: the agents produced roughly three times the work I would have managed on my own.

I will softly admit that nowadays, I don't touch files in the repo, I don't write the code, and I often don't read the code. What wound up working well for me was less coordinating threads and more - as Claude borrowed from my prompt - herding them: arranging them, sometimes just in terminal windows running in process-when-done order. And then going through a workflow of work prompts and review prompts that wound up with this kind of output, basically knocking out pre-MVP features at the rate I did in the home stretch.

Seeing the numbers quantified like this was startling even to me. Many of the features done in the last few months have been medium impact, not engine-level, but they all took about the same real effort as features that would have taken me days to write by hand.

The CLAUDE and AGENTS markdown file is actually one of the most valuable files in the repo. It took careful and pretty pointed crafting to get it where I wanted it. The goal was simple:

All agentic code created in your repos should be identical to code you would have written without it.

The real key to this level of productivity is that senior engineers are usually doing two jobs to deliver a feature: understanding the requirements and writing the code. With the code taken care of, I can just move through more requirements - which really just means more features.

Even the best models in these tools still get a lot of the code wrong, as you can see above - it's still about 50/50 whether a heavily planned feature will fail a simple lint and test-all after it says it's finished. But what they are best at, by far, is review, and even better, reviewing themselves. Once the workflow is dialed in, the code can be looked at in as much detail as you want.

What really makes this workflow work is the restriction around the data model. The first thing I do when spinning up these agents is basically full access - let's go - but I don't dare let them change things that matter, like the structure of cross-cutting schemas.

Anyone who knows exactly what they want, and can build it themselves, can harness these tools into doing it for them at at least 3x the velocity.

This is after asking it to tone it down.

P.S. After finishing this post, I asked the much less flowery codex-5.4 the same prompt. I'll leave you with its opening: