← Prev Next →

AI Presentation 2026 - LLM AI Agents

Working with an LLM from a terminal, with access to local files and tools

1.LLM AI Agents Concept Diagram

Agents are an advanced topic that you have read about in the news, but are unlikely to use anytime soon. However it is useful to understand their concept because they play a large role in the Benefits and Liabilities of LLM AIs. A user hands the agent a goal. The agent loops: build prompt and context, call the LLM, interpret the response, execute an action with tools such as web search, databases, files, or email, then update state and check the goal. It repeats until done and returns a final response.

2. Notes

  1. The main thing is to note that they run autonomusly from an initial prompt you set and a stopping goal you set.
  2. The good news is when they run well you get a lot done without your immediate attention.
  3. The bad news is that they can create a lot of junk output and delete good content when they don't operate the way you want. Sometimes that can be difficult to prevent because so much of what they do is based on inferred meaning from your request.
    So developers test in a "sandbox" until everything works then deploy to working repositories.
  4. The really bad news is that bad actors can create swarms of agents to attack banking security, civilian infrastructure control, and post miss-information on social media.