I think step 4 is the agent swarm. Manager model gets the prompt and spins up a swarm of looping subagents, maybe assigns them different approaches or subtasks, then reviews results, refines the context files and redeploys the swarm on a loop till the problem is solved or your credit card is declined.