I started at Loop on 28 September. On day two I put forward my first change to the software. Engineers call this a pull request: a proposal that a colleague reviews before it joins the product. It was accepted on day three and went out in the next release (a scheduled update of the software), in my second week. That sounds like a story about speed. It isn't, quite.
The change was small: a way to run the whole Loop software on my own computer, without needing the company's online servers. I wrote it because I needed it. The writing took an afternoon. Getting to the point where I could write it took longer, mostly getting access and setting up the right working environments, plus a few answers nobody had written down. That pattern held for the whole fortnight. The slow part of a new engineer's first weeks was never the typing. It was being unblocked.
That shaped how I used AI tools. An AI assistant can write a lot of code quickly, but an engineer who is waiting for a permission or a database gets nothing from that. So I spent the first week removing waits, for myself and then for the people around me. The second week went on working out how to point AI at the work that was left.
On 9 October a small group from Loop spent a day with Lee Goldsworthy and William Foster, who run an AI engineering workshop. Lee describes four levels of AI engineering.
At level 1, an engineer copies code into a chat tool such as claude.ai and asks for a change. Then they paste the result back into the product's code. At level 2, the engineer uses a coding assistant such as Claude Code to change the code directly. (It reads and edits the code itself.) At level 3, the engineer manages a team of AI assistants, called agents. Each agent suits a different engineering task, and the team takes larger jobs with more freedom. At level 4, the engineer directs that team inside strict safety checks. They spend the day describing what is needed and checking the results, not writing code.
Loop wants its engineers to move up these levels one at a time. I picked level 4 to test, because I wanted to find out what it takes on our own software.
The next day I started building a harness, which is the program that runs the agents and checks their work. You give it one task from our work list. It runs a team of agents through the same steps a human team follows at Loop: plan, write the code, check it, review it, fix what the review found. It ends with a proposed change for me to review, marked as not final. Getting an agent to write code was the easy part. Getting a team of them to follow our task lists, our tests, our review process and our release rhythm took the effort.
Two weeks in, this is what exists. The harness runs Claude Code with no one at the keyboard, inside a sealed box on a single computer, against a working copy of the Loop software that holds demo data. Demo data is made-up records. No experiment uses customer information. Each step has its own job, and the harness decides whether the checks passed, not the agent. That matters because the harness is ordinary code, so it is predictable: the same inputs give the same answer, every time.
The agent also writes small scripts that show the change works. The harness runs them and keeps the records and screenshots. Then the proof is checked twice. Plain code says pass or fail, and a second AI model reads the proof against what the task asked for. All of it is attached to the proposed change, so a reviewer can see what was checked and what wasn't, and an auditor can see it later.
I review it before it goes to anyone else, and a colleague reviews it again. I add it to the product through the harness, which won't let me until the automated checks and the reviews pass.
It is early and there is more to come. I'll write up what I learn roughly every two weeks, and I'd like to work it out in the open.
Glossary
Agent: an AI model that can take steps on its own, such as reading files and running commands, to finish a task.
Claude Code: Anthropic's coding assistant. It works inside an engineer's own code.
Demo data: made-up records used for testing. It holds no customer information.
Harness: the program that runs the agents, applies the checks, and decides whether a change passed.
Predictable: the same inputs always give the same answer. Engineers call this deterministic.
Pull request: a proposed change to the software that a colleague reviews before it is added. A draft pull request is one marked as not yet final.
Release: a scheduled update of the software that goes out to be used.
Sealed box: an isolated space on a computer. Software inside it cannot touch the rest of the machine. Engineers call it a container.
Tests: small automatic checks that tell you whether the software still does what it should.