Member-only story
AI Agents: Useful and Practical Proof of Concept Using Autogen
I’m surprised it worked and was useful, to be perfectly honest
I just got AI agents to do something useful as a proof of concept, using natural language to generate and execute code to search the arXiv via outputting the references into a spreadsheet and JSON file, in particular the preprints I uploaded there in physics grad school. Very surprising it worked, to be perfectly honest, lol. Here are the high level and accessible points without getting bogged down by the technical details that will follow after.
High Level Intro
- Autogen: I’m using an open source framework called Autogen in Python. Here’s the repo and paper.
- Actions: The framework can do many things. In this case, I used it to generate code (to query the arXiv for papers I’ve published), execute code (to actually do it), and assess the code (to see if the results are satisfactory). If it doesn’t meet some criteria in the output, then the loop continues on until the criteria is met or for some number of trials.
- Agents: Where do the agents fit in? The framework does many things, but this simple case will involve three types of agents: Code Generator, Code Executor, and Manager. You talk with the Manager agent, and it orchestrate the other agents as mediated by the framework, and there are API calls to OpenAI’s GPT-4o for the natural language glue between them, as well as the code generation…