Most people see "multi-agent" and think it requires building an architecture.

In reality, you just need to have the model think from three different perspectives simultaneously.
The role-playing prompt accomplishes what the agent swarm's core mechanism does: forcing the model to start from different premises, creating internal tension, and preventing conclusions from following the same line of thought.
Architectural approaches make this process explicit by breaking it into separate agents, contexts, and calls.
This is costly but highly controllable, suitable for production environments that require auditing.
However, 90% of use cases do not need auditing; they need answers that are not just one voice.
Here's an underestimated engineering fact: language models are trained on massive amounts of text from various roles.
Switching roles for the model is not simulation; it activates different weight distributions.
Give it three roles, and it will genuinely process the same material using three different cognitive structures.
The complexity isn't in the tool itself but in whether the prompt clearly states the role assumptions.
If clearly stated, one prompt equals a swarm.
If not, no matter how beautiful the architecture, you'll get three agents giving you the same answer.
---Prompt Example---
You are now playing three roles. Independently think about the following content, then discuss with each other.
Before starting, define in one sentence: what problem does this material aim to solve, and what specific result should be achieved?
All three roles must base their understanding on this definition and cannot interpret the goal differently.