Skip to content
All articles

KNOWLEDGE · Claude in practice

How much of my AI OS runs without me? 0.8 percent.

Published on

Of the 1,211 changes that landed in my AI OS over the past thirty days, 10 came in without me doing anything. That is 0.8 percent. My AI OS is the system Andy Intelligence runs on, with agents that plan, build and keep things up to date for all my roles, at night too. I wanted to know how much of that happens on its own, so I started measuring it. The figure is lower than the story I had about it.

In this article

What I measure

The figure is the share of my work that gets done without me touching it. One change is one commit on the main branch of my repository, and a merged pull request counts once. Only two routes count as without me: a pull request that merged itself, and a commit that an agent in the cloud wrote straight to the main branch. Everything else counts as with me, including a session in which an agent did the work while I sat next to it. Logs the agents write about their own runs don't count, because they change nothing in the system.

Why it is so low

Of the 1,211 changes, 1,191 came from a session I was part of. In the same thirty days, 48 pull requests merged themselves, and 42 of those changed nothing but a log. So the agents that run at night mostly write their own diary. What they propose in substance, I merge myself. If I count the logs, I get 4.2 percent, and that figure mostly measures how much gets written.

What the figure can't see

The figure only sees what lands in git. An n8n workflow that processes an email, a note in my knowledge base or a message through Telegram doesn't show up in it, even though work happens there without me too. I don't know how large that share is. The code of this site and of other websites I work on lives in separate repositories and doesn't count either.

Where it is heading

My direction is for Andy Intelligence to run mostly on agents, day and night, with me there when a decision is needed. In my head that is ninety to ninety-five percent. That is a direction. The measurement says 0.8 now. The first step is for a pull request of a safe kind to land on its own, without me pressing merge. The measurement stays the same.

What this means if you let AI work alongside you

For a company that lets AI work alongside its people, the question is the same. How much does it really do by itself? For me the definition made a difference of a factor five. With the logs it was 4.2 percent, without them 0.8. That is why I start with the question of what counts, and only then with the figure.

How it was measured

Definition
The share of changes on the main branch of my AI OS that got there without me doing anything. A merged pull request counts once. Logs and the agents' bookkeeping don't count.
Result
10 without me, 1,201 with me, 0.8 percent.
Breakdown
Of the 1,201 with me, 1,191 came from a session I was part of, 8 from a pull request I merged myself and 2 from a cloud commit that came in through a merge. The 10 without me are 6 pull requests that merged themselves and 4 commits an agent in the cloud wrote straight to the main branch. In total 48 pull requests merged themselves, and 42 of those changed nothing but a log. If I count the logs, it is 58 of 1,373, 4.2 percent.
Window
31 August to 29 September 2026.
Measured on
3 days old
Command
uv run --python 3.12 scripts/autonomy-baseline.py --ref b40e6a44c0e3eed45f13d81d985bdd0d62fdf525 --until 2026-09-29 --window-days 30
Check it yourself
My repository is private. That is why the full list of 1,416 commits is here, with date, short hash, class, rule and the number of changed paths, without subjects and without file names. Download the list (CSV)
Previous measurement
0.3 percent on 23 September 2026.

I'm curious where the next measurement lands!