What dozens of AI agents taught me: how I wrote the Blockstor storage system as an experiment

The author decided to create a clean-room implementation of LINSTOR, a distributed block storage system, using AI agents. Named Blockstor, it is an orchestrator for block devices that supports replication, snapshots, and failover. The author chose to reuse LINSTOR's API types as contracts and use TDD to define behavior, since LINSTOR does not publish its test suite. They used golangci-lint and set up a test environment with Kubernetes on Talos, requiring DRBD. The first prototype was request-based, but the author asked the AI to rewrite it using controller-runtime for asynchronous behavior. They then generated a list of issues from their own debugging articles and reports, and had the AI fix them. To speed up the process, they parallelized agents, with up to 60 agents working on separate worktrees and environments, half on unit tests and half on e2e tests. The main agent merged their fixes. After many iterations, the system became more stable, but still required manual integration testing.