How do I implement a Multi-Agent System (MAS) for complex research?

All answers to this question.

Role-based collaboration is the secret sauce for complex tasks. In a framework like CrewAI, you define a "Manager" agent or a "Sequential Process" where the Researcher first outputs a markdown file of facts. The Analyst agent then takes that specific file as input to look for trends, and finally, the Writer uses the Analyst's report to create the summary. By siloing their responsibilities and providing "Backstories," you reduce the cognitive load on each individual LLM call. This prevents the "distraction" you're seeing because each agent only focuses on a very narrow slice of the overall goal.

   Answered 2024-07-12 by Jessica Adams

What are you using for the "shared state" or blackboard? If agents don't have a central place to store and retrieve the progress of the project, they will inevitably start repeating the same search queries.

   Answered 2024-07-13 by Robert Thompson

• Robert, we were originally just passing strings between them, but that was clearly a mistake! Jessica, your mention of "Backstories" helped us realize our agents were too generic. We’ve now given our Researcher a "Skeptical Fact-Checker" persona, and it’s already producing much higher-quality data points. We are moving to a "Stateful" architecture where all agents write to a shared JSON object so they can see exactly what has been completed in real-time.

     Commented 2024-07-15 by Christopher Evans

Multi-agent systems also allow you to use different models for different roles. You can use a cheaper model for the "Researcher" and a more expensive, logical model for the "Analyst."

   Answered 2024-07-16 by Linda Moore

• I agree with Linda. This "mix and match" approach is a great way to optimize your compute budget while still getting high-reasoning results where it matters most.

     Commented 2024-07-17 by Jessica Adams