Multi-Agent Role-Play
Hand one agent the whole job and it has to be designer, implementer and critic in a single stream of text, judging work it produced moments earlier. Multi-agent role-play splits the job by giving each agent a part to play — a remit, a tool set, and a system prompt written for that part alone.
What a role actually consists of
A role is not a costume; it is a set of concrete differences:
- its own system prompt, stating what this agent is responsible for and what it must leave alone;
- its own tools and permissions — the tester can run the suite, the writer cannot deploy;
- its own context, so a reviewer sees the artefact without the author's reasoning attached;
- its own stop condition, which is what lets the next role begin.
CAMEL paired a user role with an assistant role to drive task dialogue; ChatDev and MetaGPT stage a software company, passing work from product manager to architect to engineer to tester; CrewAI makes role, goal and backstory first-class fields on an agent.
Why it matters
The benefit is mechanical, not theatrical: narrower instructions, fewer tools to choose between, and a critic with no attachment to the work under review. The failure is cosmetic casting — one model behind every persona, so the blind spots are identical, and three agents produce agreeable chatter at three times the cost. Roles earn their keep when each one has something genuinely different: different tools, different data, or a different criterion for success.
Multi-agent role-play assigns each agent a defined part — its own system prompt, tools, permissions and context — so a plan is written, executed and reviewed by different agents rather than by one. The separation yields a critic with no attachment to the work it judges. It only helps when the roles differ substantively: identical models in different costumes share every blind spot and multiply the cost.
A film gets made because someone directs, someone films, and someone else watches the footage back looking for mistakes. Nobody does all three at once, and the person checking the work is not the person who shot it.
Multi-agent role-play sets AI up the same way. Instead of one assistant doing everything, several are given parts — one plans, one carries the work out, one does nothing but look for what is wrong — each with its own instructions and its own limits on what it is allowed to touch.
It works when the parts genuinely differ. Giving the same assistant three job titles just gets you the same opinion three times over.