DevFlow

The problem

A small team was carrying a large and growing backlog for a compliance product in daily use by US energy operators. A sprint took around 25 days from start to release, and hiring was not going to close the gap quickly.

What I built

DevFlow puts AI coding agents inside the team’s normal way of working instead of beside it. Work still starts as a ticket in the backlog and still ends as a reviewed change going through the usual release pipeline. What changes is who does the first pass.

1. Ticket. Work starts from the team’s normal backlog. Nothing changes for product owners.
2. Plan and build. AI agents plan the change and implement it, following the project’s own conventions.
3. Review and test. Automated reviews and browser tests check the change before a person looks at it.
4. Engineer approval. An engineer reviews the change and decides. Nothing merges without that approval.
5. Release. Approved changes go out through the same pipeline and quality gates as code written by hand.

Guardrails

The design question was never “can an agent write this code”. It was “who is accountable when it does”. Every change is approved by an engineer before it merges, agents work to the project’s own rules rather than their own habits, and people only review work that has already passed its checks.

Outcome

Sprint cycles went from about 25 days to 10–11 days, and a team of three engineers now sustains the output that previously needed six. DevFlow is in daily production use.

My role

I conceived DevFlow, designed it and led its build and rollout. I also trained the team to work with and supervise AI coding agents.