Delete my own platform
Governed AI operations, and why I removed most of what I had built to run them.
Context
I built an AI DevOps agent with one rule: the agent may propose any infrastructure change, and an approval on record is the control. Every change is a pull request, and every action lands in a tamper-evident audit trail.
It worked. The trail holds 3,948 audited actions since July 2026. Then I measured the platform against its real use: about 390K lines across five repositories, and one tool made 93% of the agent’s changes.
Approval to execution was measured at about two minutes, and a 90-second hold brake was proven on a real change. In one 30-day window, 43% of decisions were made by policy, 27% on the code host and 30% by a person.
Decision
Simplify. Replace the custom agent with a managed agent, move approvals into rules the code host already enforces, and retire the custom gateway.
What we did
Non-production changes are approved by a written policy, and each automatic approval is recorded as a policy decision. Access grants and production changes stay with a person.
Before the cutover I compared the old and new paths: 388 of 389 action records were identical. The gateway was retired on 10 November 2026 after a zero-traffic window, with a one-command rollback kept.
Result
In the eight weeks to 15 November 2026 the agent proposed 338 changes across 21 services and 5 environments, with a median of 4 minutes from proposal to merge. Each was approved by a person or the written policy. None touched production.
Lesson
Measure what you built against how it is used, and be willing to delete it. The control that mattered was the approval on record, not the platform around it.