Onyx Edge
Request a consultation

Case study

AI agent delivery system

The system I built to run software delivery with AI coding agents: every change is planned, tested, reviewed, and recorded, so the work can be audited long after the session that produced it.

  • 2025 to now
  • Designed, built, and operated
  • Go, Kubernetes, and MCP

The problem

AI coding agents can produce more change in a day than a team can review in a week. Without structure, nobody can say what was changed, why, or whether it was checked, and the reasoning disappears when the session ends.

The approach

An AI-augmented delivery workflow

  1. Plan
  2. Build
  3. Test
  4. Review
  5. Record
Every change planned, tested, reviewed, and recorded. It is how I build my own products.
  1. Plan

    Every piece of work is an item on a board, grouped into plans and projects, with a written description of what done means.

  2. Test first

    Tests come first: a failing test, then the change, then the full suite before anything is marked done.

  3. Decide

    Strategic choices are recorded as decisions with their context, evidence, alternatives, and success criteria. A decision is never edited; a new one supersedes it.

  4. Run unattended

    Unattended runs execute as Kubernetes jobs on their own branch and are handed back for review instead of merging themselves.

  5. Promote

    Changes move from development to staging on their own, and to production only as an explicit, separate step.

The outcome

3,400+
Items closed on its own board
16k+
GitHub contributions in the last year, many through my AI agent delivery system
  • The same process runs the Visiting Objects platform and this site, so every change on them traces back to an item and a recorded reason.
  • Months later, the question of why something was built a certain way has a written answer.

Bring the same discipline to your systems

Tell me what you build and run, and I will scope how to plan, test and review it this way.