Lifetime Citizen Portal Access — AI Briefings, Alerts & Unlimited Follows
eWorld presents agentic AI and Model Context Protocol to computer science students
Loading...
Summary
Yifan Zhao of contractor eWorld Enterprise Solutions gave a virtual workshop explaining agentic AI, retrieval-augmented generation and the Model Context Protocol, and discussed uses ranging from travel planning agents to automated customer service and state benefit systems.
Yifan Zhao, lead data scientist at eWorld Enterprise Solutions, introduced agentic artificial intelligence and the Model Context Protocol during a virtual workshop for computer science undergraduates.
Zhao framed agentic AI as systems that go beyond standalone large language models by autonomously planning, choosing tools and executing tasks; he emphasized Model Context Protocol (MCP) as an open standard that simplifies connecting AI models to external data and APIs.
Zhao said, “agentic AI is AI systems that can autonomously plan, make decisions and execute tasks across multiple systems to achieve specific goals.” He described retrieval-augmented generation (RAG) as a method to reduce hallucinations by retrieving verified documents and feeding those as supporting context to a generative model.
The presentation traced generative AI’s roots in natural language processing and the transformer architecture. Zhao summarized key technical ideas in plain terms for the student audience: the transformer’s attention mechanism, autoregressive next-token generation used by GPT-family models, and the practical limits—most notably hallucination and the engineering complexity of reliably connecting models to external tools.
Using concrete examples, Zhao showed how an agent can combine web search, multiple language models and diagram APIs to produce a text itinerary plus a route diagram. He said diagram generation typically requires calling a diagram-drawing service such as Lucidchart or draw.io, noting that “the large language model is still a language model” and often must call other services to produce nontext outputs.
On MCP, Zhao described the protocol as “an open standard that [is] designed to allow AI models to seamlessly access external context from various sources such as databases, files and APIs.” He said major providers—including OpenAI, Anthropic, Google, Meta and Amazon—have MCP support and that the protocol reduces integration work by translating requests and structuring responses between models and tools.
Zhao recommended Google Cloud’s agent development kit as a starting place for students learning to build agents and said Google Cloud offers documentation and academic credits that may help undergraduates experiment without large up-front costs.
Selma, the workshop moderator, told students the coming hackathon season will be AI-focused and encouraged attendees to use available vendor resources. An attendee who identified herself as Angie asked specifically about coursework and learning resources; Zhao recommended cloud provider SDKs and tutorials and offered to continue the conversation in a shared Slack channel.
Zhao also described eWorld’s work with the State of Hawaii, identifying the Department of Human Services as a client and saying the company uses automation and AI to streamline benefit application, verification and delivery processes. He framed that work as motivation for students interested in applied AI in public services.
The workshop concluded with an offer to continue the discussion via Slack and other vendor channels and a reminder that MCP and agent frameworks are intended to reduce redundant engineering when integrating models with tools.
