This is from the guy that invented the infamous Ralph loop. I was browsing around looking at agent team systems and found this. Of all the ones I came across this seemed relatively pragmatic: define a pipeline of prompts/agents each with a particular role, tasks fall through the pipeline until it either concludes it is done or until a retry count is exhausted.
For example: you define a planner that breaks a issues into sub-tasks (commits). That plan is passed to a setup agent that creates a branch/dev env. Then the developer agent completes all the sub-tasks with tests, a verifier agent checks the developer agent's work, a tester agent performs E2E tests, another agent creates a PR, and another does code review.