Review
The code is reviewed by whoever did not write it
QA and security are roles separate from the executor, and they run after it. It is the same reason a business does not let one person both record and check the till.
H&I Digital does not ask an AI to build a system.
We direct a team of agents with separate roles, and we published the tool that does it.
agent-team is an open-source skill, MIT licensed, on the H&I GitHub. It turns the assistant into a team: each role has one responsibility, the executor works isolated in its own Git worktree, and decisions are recorded so the next task does not ask the same question. It is what runs behind the systems in the portfolio.
Nine roles, five phases
Scroll the page: the pipeline runs. Scrolling back rewinds it.
Phases 1 and 4 run in parallel, planning and architecture come out together, and later testing, security, infrastructure and compliance do too. Splitting by role beats a single agent because each arrives with a different criterion: whoever writes the code is not whoever hunts for the hole in it.
Pipeline runningphase 1/5
Visualisation unavailable on this device, the phases are listed below.
- P 1
Planner
Gathers the requirement and writes down what must exist.
- P 1
Architect
Designs the solution and settles the structure before any code.
- P 2
UX
Defines flow and interface from the requirement.
- P 3
Executor
Writes the code, isolated in its own Git worktree.
- P 4
QA
Tests what was built and reports back what fails.
- P 4
Security
Hunts for holes: injection, sessions, permissions, leaked secrets.
- P 4
Infra
Handles deployment, containers and what runs in production.
- P 4
Compliance
Checks data protection, retention and personal data handling.
- P 5
Context Steward
Keeps the decisions so the next task does not ask the same question.
Why this matters to you
You are not buying AI. You are buying the process built around it.
Isolation
Parallel work without one stepping on the other
Each executor works in its own git worktree. Two fronts advance at once without conflict, and whatever does not hold up is discarded without contaminating the rest.
Memory
The decision stays on record
The Context Steward keeps why each choice was made. Six months later, when you ask why the system does it one way and not another, the answer exists in writing.
Proof
It is public, you can check
The code is on GitHub under an MIT licence. It is not a methodology slide: it is a tool anyone can download, read and use.
Want to see this applied to your case?
Thirty minutes, no strings. You explain how the operation works today and we tell you what can come off people’s hands, and what is not worth automating.