Large Language Models (LLMs) use Neural Networks to identify and learn patterns from large amounts of text documents. Simply put, they read so much text, that whenever you give them the beginning of a sentence, they have a very high chance to predict how it continues.

This ability of predicting the next word for a given sentence has opened the door to interacting with machines using natural language, a whole new interface.

What does it take to build a system where multiple agents can interact asynchronously, respond to user queries and take proactive action when things change around them?

Many articles already explain how to build an agent with existing libraries and SDKs. Here, I’ll cover what else you need to create a multi-agent system: a future-proof architecture that supports asynchronous communication, works with any tooling or models, and runs locally on varied hardware.

Prerequisites This is a dense article that introduces many concepts without fully defining them. Some familiarity with LLMs, the concept of agentic AI, and how to build agents is advised, in particular with LangChain4J and Jlama.

TextRank algorithm was introduced by Rada Mihalcea and Paul Tarau in their paper “TextRank: Bringing Order into Texts” in 2004. It applies the same principle that Google’s PageRank used to discover relevant web pages.

The idea is to split a text into sentences, and then calculate a score for each sentence in terms of its similarity to the other sentences. TextRank treats sentences having common words as a link between them (like hyperlinks between web pages). Then, it applies a weight to that link based on how many words the sentences have in common. ts-textrank uses Sorensen-Dice Similarity for this.