When designing a multi-agent system for process automation, such as in a contact center, the process typically starts with analyzing historical data (e.g., call transcripts, user interactions) to understand how human agents complete tasks.
An LLM-based model is often used to analyze this data and identify the common steps a person takes. For example, looking at calls from the past six months and understanding how agents authenticate users, determine intent, etc. Since there is variability in execution, a probabilistic model helps map out these actions in a structured way.
Once a granular workflow is defined, SMEs validate and refine it, creating a detailed blueprint of human activities. This blueprint is then used to design agent milestones.
First, deciding whether workflows require one agent or a sequence of agents.
If the workflow includes multiple tasks, choose an architecture with a sequence of agents. Each agent should be assigned a specific task that can perform well to optimize efficiency—e.g., one for user authentication, another for sentiment analysis, and another for breaking down complex requests.
Finally, an orchestrator integrates these agents into a seamless system.
