What's the right way to orchestrate multiple specialised agents that need to collaborate?

We've broken a complex workflow into specialised agents: planner, retriever, coder, critic. The problem is handoffs — context is lost between agents, errors in one agent compound in the next, and debugging a multi-agent trace is a nightmare. Are teams using a shared memory bus, a supervisor model, or something else? How do you decide when to use one general-purpose agent vs. a pipeline of specialists?