Augmented LLM

An augmented LLM is a single model call extended with reaches into the world. A bare LLM (large language model) takes text and returns text: it cannot run code, look up a document, or remember what happened last time. The augmented call adds three reaches alongside the prompt: it can call a tool, pull a document from a retrieval system, or read and write memory. It is still one call, one request and one response, acting on more than the words in front of it.

This is the atomic unit. Stack and orchestrate augmented calls and you get workflows and agents; there is nothing else underneath.

Most products sold as “AI apps” are one augmented call with a good prompt, and one call is usually the right amount of machinery. Scope a build from here and add structure only when a single augmented call cannot carry the job.