What is the relationship between dynamic workflow and multi-agent, subagent, goal, and skill?

The official article of Opus 4.8 says that Claude can plan first, then schedule hundreds of parallel subagents in a single session, finally perform verification, and then return to the user. The bottleneck of many real engineering tasks is that single-threaded context processing is too slow and too fragile. For example, full warehouse bug hunt, migration of hundreds of files, cross-module security audit, and independent review of the same solution. These tasks are not completed by an agent in one go. They are more like a temporarily formed project team.