YUNOVA · LLM Fundamentals
Module 13: Agents, safety and multimodal
AI agents and tool-use loops
Key points
- An agent is a model that calls tools in a loop to complete a task.
- Each step: the model picks a tool, the tool runs, and the result goes back to the model.
- The loop ends when the model gives a final answer or hits a step limit.
Model Context Protocol (MCP)
Key points
- MCP is an open protocol for connecting AI applications to tools and data.
- An MCP server exposes tools, resources and prompts; a client connects to it.
- One server can be reused by any MCP-compatible application.
LLM guardrails and red teaming
Key points
- Guardrails check inputs and outputs to block unsafe or off-topic content.
- Prompt injection hides instructions in data to change the model's behaviour.
- Red teaming tests a system by deliberately trying to make it fail.
Multimodal models: images and audio
Key points
- Multimodal models accept more than text, such as images and audio.
- Images are split into patches and turned into tokens the model can read.
- Uses include describing images, reading documents and transcribing speech.