Traditional AI/ LLM concepts that still matter in the age of Agentic AI

Reading Time: < 1 minute

Vector Embeddings

Concept: embeddings are used for semantic (similar meaning) search

Current Utility: Still relevant in Agentic AI, for example in semantic prompt caching to reduce AI agent costs. 

Qunatization

Concept: technique used to compress large models for deploying models on constrained infrastructure 

Utility: important technique to reduce the memory and storage requirements of models in agentic workflows

Prompt Injection

Concept: an attack technique to manipulate LLMs behaviour

Utility: Still a major security risk, especially when agents can access data or external systems. 

ReAct – 

Concept: a prompt framework that combines reasoning with actions 

Utility: The foundational pattern for building tool-using AI Agents 

RAG

Concept: Retrieving relevant information stored in a vector databse at inference time and providing it to the LLM as context 

Utility: RAG pipelines can be converted into Agentic RAG systems

Open source vs open weight models

Open-weight: The trained model weights are publicly available, but other components such as training data, training code, or methodology may not be.

Open-source AI: The models are made available to use and customise (under licenses). 

Utility: Important to consider when choosing the underlying model, as it can affect how an agentic application is built.