Case study · platform engineering with AI agents

One engineer, twelve agents, ninety-one merge requests

The security team turned on a centrally enforced scanning pipeline for every repository. About a hundred repos needed a CI change before it could scan them properly, spread across fourteen teams. I ran the migration with a crew of AI agents, small background scripts and one rule: nothing reached a team without my review.

The whole approach in under a minute: the five layers, one repo going through the loop, the human gate, and the scale.
1orchestrator session
12AI agents
51background script jobs
96repos audited
91merge requests

The approach

  1. Map before fixing. Read-only agents worked out who owns each repo and where that team reviews code, from commits and tickets rather than folder paths.
  2. One agent per team group. Seven agents each took one team's repos and a shared playbook, two running at a time, instead of one agent per repo.
  3. Rules that learn. Every repo that surprised us became a new playbook rule, so later agents started smarter than earlier ones.
  4. Prove it on real pipelines. Scripts watched every pipeline and compared the image the build pushed with the image the scanner was told about.
  5. A human gate. I reviewed each team's changes and approved every message before it went out, one team at a time.

Explore it

Who did whatThe agentsOne card per agent: its instructions, what it got, what it did, what it gave, and the rules added along the way.Open the roster → How work movedThe hand-offsEvery output landing as someone else's input, in order, across eight lanes and five phases.Open the hand-offs → The whole systemThe flow mapThe system map, how each repo's fix was chosen, and the pains we hit phase by phase.Open the flow map →

Built from the migration's own records. Company, team, repository, channel and people names are changed or left out; the structure, the numbers and the lessons are real.