About This Session
Millions of dollars were invested in your company's Gen AI chatbot solution. After using it for a week, users discovered that it was able to do nothing more than respond to queries and reverted back to their usual way of doing things. This is not a problem with the chatbot, but rather the architecture itself. Using an LLM wrapped up in a text box and calling it an "enterprise AI solution" would be like making an HTML page and calling it a "web application."
In this session, we will be exploring the architectural patterns that are needed for this transition. Agent decomposition patterns are discussed, which include the decomposition process of a monolithic "do everything" agent into smaller specialized agents that are independently testable and deployable. Patterns for orchestration vs choreography are explored, discussing when it is necessary to have a controlling agent and when collaboration between agents leads to better performance through events. The topic of tool-use patterns is thoroughly explored, covering the concept of MCP (Model Context Protocol) patterns, whereby the agents detect and invoke external tools and services. The most difficult part is also addressed – the observability/governance stack for running agent-based solutions in production since when a group of agents made an autonomous decision affecting your customers, you have to justify your action.
Participants will receive the architectural plan needed to transition from chatbot to agential orchestration, drawing on established enterprise architecture principles.