Liam Williams
The right problem, solved right.
I set up AI agents that ship real deliverables in a team's codebase, while the team keeps the approval of every merge. CI/AI/CD is the name I gave that setup in March 2026, when I noticed that CI/CD has a human-shaped hole in the middle and wondered what would happen if agents lived in the gap. Eight months on, the agents live there, and so does a person, on purpose.
Since March 2026 I've used a fleet of AI coding agents to accelerate development of a commercial Java product with paying customers. From 13 March to 28 October 2026: 3,085 pull requests merged, 97% of them written by agents, seven customer releases, and every merge approved by a named person.
Lesson one was mine: my first count of their throughput was wrong by 2.4 times because I was measuring the wrong metric for counting PR merges. Other lessons (so far) are about where the constraint moves (the build, then the budget, then the person approving) and what it takes to keep a person in the loop without making them the bottleneck: a supervisor agent, rules turned into checks that refuse, a budget paced to the day, and measurement you can trust. The rest are in the article, on LinkedIn from 10 November.
Most of what's published about autonomous AI agents in software delivery points the other way, towards the "dark factory" where nobody reads the code. I've made the opposite bet, and I have eight months of numbers on it, failures included.
Before agents, I spent 15+ years in enterprise Java, microservices, test-driven development and DevOps, with 20% to 50% improvements in DORA metrics on enterprise projects. The agents ship well only because a pipeline built to that standard catches their mistakes in time. In 2017 I co-founded Traffic Parrot, a service virtualization product used by Fortune 500 companies, and now by their AI coding agents too.
First-Class MSci, Mathematics and Computer Science, Imperial College London. Distinction Postgraduate Diploma in Law, BPP Law School. Technical editor of 150+ Baeldung articles. Published on InfoQ.
Elsewhere: Traffic Parrot (service virtualization, which I co-founded), whatproblem.dev (the first question I ask: what problem are you trying to solve?), ads.txt Guru, and free guides for UK and US taxpayers taking a tax dispute to tribunal or court without a lawyer (taxtribunalhelp.co.uk and taxcourthelp.com).