CI/AI/CD

AI agents that ship real deliverables in your codebase, with a person approving every merge.

Continuous integration tells you what's broken. Continuous delivery ships what's ready. Between them sits a person. CI/AI/CD puts agents in that gap, and keeps the person there on purpose.

Supervised agentic delivery, one codebase at a time, for companies with 20 to 300 developers. By Liam Williams.

Take the scorecard

Who it's for

CTOs and VPs of Engineering at companies with about 20 to 300 developers. You want AI agents shipping real deliverables in your delivery pipeline, but you don't want to hand the merge button to a machine.

The proof

Since March 2026 I've run a fleet of AI agents on a commercial Java product with paying customers. From 13 March to 28 October 2026:

3,085pull requests merged, 97% of them written by agents

7customer releases, roughly monthly since May

3,058 to 11,240JUnit tests, March to October

106 vs 4new features in the release notes: those seven releases against the three of the six months before

zerocode commits of my own since 1 September

every oneof those merges approved by a named person

The full story, with what went wrong, is the article I publish on LinkedIn on 10 November 2026: "The human gate is the point". My first count of that throughput was wrong by 2.4 times, and I'm glad I never published it. Lesson one.

What I set up with your team

Agents that work your backlog. They pick an item, write the code and its tests, build, raise a pull request and answer review. They run in containers you control, with no production credentials, and reach only the hosts you approve.
A supervisor. It reviews the agents' work and settles anything a revert can undo. It brings your people only what needs them.
A human gate that stays cheap. Every change is reviewed by another agent, then by the supervisor. Your approvers get ready pull requests in batches. A product question reaches them as a single question, not buried in a review.
Hard rails. Your rules become checks that refuse a change, not documents an agent can forget.
Pacing and measurement from day one. You see what it costs, what it ships and what fails.

How we work together

Readiness scorecard (free)
Eight questions about your build, backlog, review capacity, secrets, budget, supervision, rails and measurement. Then a 30-minute call on your answers.
Discovery assessment (2 weeks)
I review your pipeline and codebase against the scorecard. You get a readiness report, a pilot plan with its first pieces of work, the design of your human gate, and a cost model.
Pilot (4 weeks)
Agents run on one real codebase of yours, with one named approver and a stand-in, and with the supervisor, rails, pacing and measurement in place. At the end, I hand it to your team with the numbers.
Retainer (monthly, optional)
Tuning, new rails, and a review of the numbers as you scale.

What it isn't

Not a "dark factory". Your people keep the approval.
Not a slide deck. The pilot ends with agents merging real deliverables under your rules.
Not a promised number. The scorecard says first whether your pipeline is ready. Two of its answers can end the conversation, and if they do, I'll say so rather than sell you a pilot.
Not a rollout. A pilot is one codebase and one named approver: the size eight months proved. Wider is planned from what the pilot shows, not promised.

Next step

Take the scorecard Message me on LinkedIn

Say it like CI/CD, with the AI in the gap.