What it does
The pipeline gathers signals, scores each opportunity and puts a ranked list in front of the people who act on it. The output is a worklist inside the tool the team already uses, not a chat window they have to question.
How it is governed
- A typed registry. Every agent is declared in one place with a named owner, an autonomy tier and a cost ceiling. If nobody owns it, it does not run.
- Read first, write later. Most agents read and suggest. Anything that writes to a business system needs a person to approve it.
- Evaluations before production. The pass threshold is written down before the test runs, so it cannot move afterwards.
- Gates set to measured reality. Coverage floors start at the level the code actually has and only go up. A gate set above reality gets switched off, and then it protects nothing.
- Suspect data stays visible. A value that looks wrong is flagged, not replaced with a better-looking guess. Commercial judgements go back to the business owner.
My part
I designed the pipeline and wrote its first versions, the CI and coverage gates, and the rules for containing bad data. Engineers on my team built much of what runs today, and they own it.
What I would keep next time
- One agent, one owner, one measurable process.
- Read-only until it has earned the right to write.
- Evaluations and thresholds agreed before the build.
- A date by which it is either in use or switched off.